LLVM API Documentation
00001 //===-- llvm/Target/TargetLowering.h - Target Lowering Info -----*- C++ -*-===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 /// 00010 /// \file 00011 /// This file describes how to lower LLVM code to machine code. This has two 00012 /// main components: 00013 /// 00014 /// 1. Which ValueTypes are natively supported by the target. 00015 /// 2. Which operations are supported for supported ValueTypes. 00016 /// 3. Cost thresholds for alternative implementations of certain operations. 00017 /// 00018 /// In addition it has a few other components, like information about FP 00019 /// immediates. 00020 /// 00021 //===----------------------------------------------------------------------===// 00022 00023 #ifndef LLVM_TARGET_TARGETLOWERING_H 00024 #define LLVM_TARGET_TARGETLOWERING_H 00025 00026 #include "llvm/ADT/DenseMap.h" 00027 #include "llvm/CodeGen/DAGCombine.h" 00028 #include "llvm/CodeGen/RuntimeLibcalls.h" 00029 #include "llvm/CodeGen/SelectionDAGNodes.h" 00030 #include "llvm/IR/Attributes.h" 00031 #include "llvm/IR/CallSite.h" 00032 #include "llvm/IR/CallingConv.h" 00033 #include "llvm/IR/InlineAsm.h" 00034 #include "llvm/IR/Instructions.h" 00035 #include "llvm/IR/IRBuilder.h" 00036 #include "llvm/MC/MCRegisterInfo.h" 00037 #include "llvm/Target/TargetCallingConv.h" 00038 #include "llvm/Target/TargetMachine.h" 00039 #include <climits> 00040 #include <map> 00041 #include <vector> 00042 00043 namespace llvm { 00044 class CallInst; 00045 class CCState; 00046 class FastISel; 00047 class FunctionLoweringInfo; 00048 class ImmutableCallSite; 00049 class IntrinsicInst; 00050 class MachineBasicBlock; 00051 class MachineFunction; 00052 class MachineInstr; 00053 class MachineJumpTableInfo; 00054 class Mangler; 00055 class MCContext; 00056 class MCExpr; 00057 class MCSymbol; 00058 template<typename T> class SmallVectorImpl; 00059 class DataLayout; 00060 class TargetRegisterClass; 00061 class TargetLibraryInfo; 00062 class TargetLoweringObjectFile; 00063 class Value; 00064 00065 namespace Sched { 00066 enum Preference { 00067 None, // No preference 00068 Source, // Follow source order. 00069 RegPressure, // Scheduling for lowest register pressure. 00070 Hybrid, // Scheduling for both latency and register pressure. 00071 ILP, // Scheduling for ILP in low register pressure mode. 00072 VLIW // Scheduling for VLIW targets. 00073 }; 00074 } 00075 00076 /// This base class for TargetLowering contains the SelectionDAG-independent 00077 /// parts that can be used from the rest of CodeGen. 00078 class TargetLoweringBase { 00079 TargetLoweringBase(const TargetLoweringBase&) LLVM_DELETED_FUNCTION; 00080 void operator=(const TargetLoweringBase&) LLVM_DELETED_FUNCTION; 00081 00082 public: 00083 /// This enum indicates whether operations are valid for a target, and if not, 00084 /// what action should be used to make them valid. 00085 enum LegalizeAction { 00086 Legal, // The target natively supports this operation. 00087 Promote, // This operation should be executed in a larger type. 00088 Expand, // Try to expand this to other ops, otherwise use a libcall. 00089 Custom // Use the LowerOperation hook to implement custom lowering. 00090 }; 00091 00092 /// This enum indicates whether a types are legal for a target, and if not, 00093 /// what action should be used to make them valid. 00094 enum LegalizeTypeAction { 00095 TypeLegal, // The target natively supports this type. 00096 TypePromoteInteger, // Replace this integer with a larger one. 00097 TypeExpandInteger, // Split this integer into two of half the size. 00098 TypeSoftenFloat, // Convert this float to a same size integer type. 00099 TypeExpandFloat, // Split this float into two of half the size. 00100 TypeScalarizeVector, // Replace this one-element vector with its element. 00101 TypeSplitVector, // Split this vector into two of half the size. 00102 TypeWidenVector // This vector should be widened into a larger vector. 00103 }; 00104 00105 /// LegalizeKind holds the legalization kind that needs to happen to EVT 00106 /// in order to type-legalize it. 00107 typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind; 00108 00109 /// Enum that describes how the target represents true/false values. 00110 enum BooleanContent { 00111 UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage. 00112 ZeroOrOneBooleanContent, // All bits zero except for bit 0. 00113 ZeroOrNegativeOneBooleanContent // All bits equal to bit 0. 00114 }; 00115 00116 /// Enum that describes what type of support for selects the target has. 00117 enum SelectSupportKind { 00118 ScalarValSelect, // The target supports scalar selects (ex: cmov). 00119 ScalarCondVectorVal, // The target supports selects with a scalar condition 00120 // and vector values (ex: cmov). 00121 VectorMaskSelect // The target supports vector selects with a vector 00122 // mask (ex: x86 blends). 00123 }; 00124 00125 static ISD::NodeType getExtendForContent(BooleanContent Content) { 00126 switch (Content) { 00127 case UndefinedBooleanContent: 00128 // Extend by adding rubbish bits. 00129 return ISD::ANY_EXTEND; 00130 case ZeroOrOneBooleanContent: 00131 // Extend by adding zero bits. 00132 return ISD::ZERO_EXTEND; 00133 case ZeroOrNegativeOneBooleanContent: 00134 // Extend by copying the sign bit. 00135 return ISD::SIGN_EXTEND; 00136 } 00137 llvm_unreachable("Invalid content kind"); 00138 } 00139 00140 /// NOTE: The constructor takes ownership of TLOF. 00141 explicit TargetLoweringBase(const TargetMachine &TM, 00142 const TargetLoweringObjectFile *TLOF); 00143 virtual ~TargetLoweringBase(); 00144 00145 protected: 00146 /// \brief Initialize all of the actions to default values. 00147 void initActions(); 00148 00149 public: 00150 const TargetMachine &getTargetMachine() const { return TM; } 00151 const DataLayout *getDataLayout() const { return DL; } 00152 const TargetLoweringObjectFile &getObjFileLowering() const { return TLOF; } 00153 00154 bool isBigEndian() const { return !IsLittleEndian; } 00155 bool isLittleEndian() const { return IsLittleEndian; } 00156 00157 /// Return the pointer type for the given address space, defaults to 00158 /// the pointer type from the data layout. 00159 /// FIXME: The default needs to be removed once all the code is updated. 00160 virtual MVT getPointerTy(uint32_t /*AS*/ = 0) const; 00161 unsigned getPointerSizeInBits(uint32_t AS = 0) const; 00162 unsigned getPointerTypeSizeInBits(Type *Ty) const; 00163 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const; 00164 00165 EVT getShiftAmountTy(EVT LHSTy) const; 00166 00167 /// Returns the type to be used for the index operand of: 00168 /// ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT, 00169 /// ISD::INSERT_SUBVECTOR, and ISD::EXTRACT_SUBVECTOR 00170 virtual MVT getVectorIdxTy() const { 00171 return getPointerTy(); 00172 } 00173 00174 /// Return true if the select operation is expensive for this target. 00175 bool isSelectExpensive() const { return SelectIsExpensive; } 00176 00177 virtual bool isSelectSupported(SelectSupportKind /*kind*/) const { 00178 return true; 00179 } 00180 00181 /// Return true if multiple condition registers are available. 00182 bool hasMultipleConditionRegisters() const { 00183 return HasMultipleConditionRegisters; 00184 } 00185 00186 /// Return true if the target has BitExtract instructions. 00187 bool hasExtractBitsInsn() const { return HasExtractBitsInsn; } 00188 00189 /// Return the preferred vector type legalization action. 00190 virtual TargetLoweringBase::LegalizeTypeAction 00191 getPreferredVectorAction(EVT VT) const { 00192 // The default action for one element vectors is to scalarize 00193 if (VT.getVectorNumElements() == 1) 00194 return TypeScalarizeVector; 00195 // The default action for other vectors is to promote 00196 return TypePromoteInteger; 00197 } 00198 00199 // There are two general methods for expanding a BUILD_VECTOR node: 00200 // 1. Use SCALAR_TO_VECTOR on the defined scalar values and then shuffle 00201 // them together. 00202 // 2. Build the vector on the stack and then load it. 00203 // If this function returns true, then method (1) will be used, subject to 00204 // the constraint that all of the necessary shuffles are legal (as determined 00205 // by isShuffleMaskLegal). If this function returns false, then method (2) is 00206 // always used. The vector type, and the number of defined values, are 00207 // provided. 00208 virtual bool 00209 shouldExpandBuildVectorWithShuffles(EVT /* VT */, 00210 unsigned DefinedValues) const { 00211 return DefinedValues < 3; 00212 } 00213 00214 /// Return true if integer divide is usually cheaper than a sequence of 00215 /// several shifts, adds, and multiplies for this target. 00216 bool isIntDivCheap() const { return IntDivIsCheap; } 00217 00218 /// Returns true if target has indicated at least one type should be bypassed. 00219 bool isSlowDivBypassed() const { return !BypassSlowDivWidths.empty(); } 00220 00221 /// Returns map of slow types for division or remainder with corresponding 00222 /// fast types 00223 const DenseMap<unsigned int, unsigned int> &getBypassSlowDivWidths() const { 00224 return BypassSlowDivWidths; 00225 } 00226 00227 /// Return true if pow2 sdiv is cheaper than a chain of sra/srl/add/sra. 00228 bool isPow2SDivCheap() const { return Pow2SDivIsCheap; } 00229 00230 /// Return true if Flow Control is an expensive operation that should be 00231 /// avoided. 00232 bool isJumpExpensive() const { return JumpIsExpensive; } 00233 00234 /// Return true if selects are only cheaper than branches if the branch is 00235 /// unlikely to be predicted right. 00236 bool isPredictableSelectExpensive() const { 00237 return PredictableSelectIsExpensive; 00238 } 00239 00240 /// isLoadBitCastBeneficial() - Return true if the following transform 00241 /// is beneficial. 00242 /// fold (conv (load x)) -> (load (conv*)x) 00243 /// On architectures that don't natively support some vector loads efficiently, 00244 /// casting the load to a smaller vector of larger types and loading 00245 /// is more efficient, however, this can be undone by optimizations in 00246 /// dag combiner. 00247 virtual bool isLoadBitCastBeneficial(EVT /* Load */, EVT /* Bitcast */) const { 00248 return true; 00249 } 00250 00251 /// \brief Return if the target supports combining a 00252 /// chain like: 00253 /// \code 00254 /// %andResult = and %val1, #imm-with-one-bit-set; 00255 /// %icmpResult = icmp %andResult, 0 00256 /// br i1 %icmpResult, label %dest1, label %dest2 00257 /// \endcode 00258 /// into a single machine instruction of a form like: 00259 /// \code 00260 /// brOnBitSet %register, #bitNumber, dest 00261 /// \endcode 00262 bool isMaskAndBranchFoldingLegal() const { 00263 return MaskAndBranchFoldingIsLegal; 00264 } 00265 00266 /// Return true if target supports floating point exceptions. 00267 bool hasFloatingPointExceptions() const { 00268 return HasFloatingPointExceptions; 00269 } 00270 00271 /// Return the ValueType of the result of SETCC operations. Also used to 00272 /// obtain the target's preferred type for the condition operand of SELECT and 00273 /// BRCOND nodes. In the case of BRCOND the argument passed is MVT::Other 00274 /// since there are no other operands to get a type hint from. 00275 virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const; 00276 00277 /// Return the ValueType for comparison libcalls. Comparions libcalls include 00278 /// floating point comparion calls, and Ordered/Unordered check calls on 00279 /// floating point numbers. 00280 virtual 00281 MVT::SimpleValueType getCmpLibcallReturnType() const; 00282 00283 /// For targets without i1 registers, this gives the nature of the high-bits 00284 /// of boolean values held in types wider than i1. 00285 /// 00286 /// "Boolean values" are special true/false values produced by nodes like 00287 /// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND. 00288 /// Not to be confused with general values promoted from i1. Some cpus 00289 /// distinguish between vectors of boolean and scalars; the isVec parameter 00290 /// selects between the two kinds. For example on X86 a scalar boolean should 00291 /// be zero extended from i1, while the elements of a vector of booleans 00292 /// should be sign extended from i1. 00293 /// 00294 /// Some cpus also treat floating point types the same way as they treat 00295 /// vectors instead of the way they treat scalars. 00296 BooleanContent getBooleanContents(bool isVec, bool isFloat) const { 00297 if (isVec) 00298 return BooleanVectorContents; 00299 return isFloat ? BooleanFloatContents : BooleanContents; 00300 } 00301 00302 BooleanContent getBooleanContents(EVT Type) const { 00303 return getBooleanContents(Type.isVector(), Type.isFloatingPoint()); 00304 } 00305 00306 /// Return target scheduling preference. 00307 Sched::Preference getSchedulingPreference() const { 00308 return SchedPreferenceInfo; 00309 } 00310 00311 /// Some scheduler, e.g. hybrid, can switch to different scheduling heuristics 00312 /// for different nodes. This function returns the preference (or none) for 00313 /// the given node. 00314 virtual Sched::Preference getSchedulingPreference(SDNode *) const { 00315 return Sched::None; 00316 } 00317 00318 /// Return the register class that should be used for the specified value 00319 /// type. 00320 virtual const TargetRegisterClass *getRegClassFor(MVT VT) const { 00321 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 00322 assert(RC && "This value type is not natively supported!"); 00323 return RC; 00324 } 00325 00326 /// Return the 'representative' register class for the specified value 00327 /// type. 00328 /// 00329 /// The 'representative' register class is the largest legal super-reg 00330 /// register class for the register class of the value type. For example, on 00331 /// i386 the rep register class for i8, i16, and i32 are GR32; while the rep 00332 /// register class is GR64 on x86_64. 00333 virtual const TargetRegisterClass *getRepRegClassFor(MVT VT) const { 00334 const TargetRegisterClass *RC = RepRegClassForVT[VT.SimpleTy]; 00335 return RC; 00336 } 00337 00338 /// Return the cost of the 'representative' register class for the specified 00339 /// value type. 00340 virtual uint8_t getRepRegClassCostFor(MVT VT) const { 00341 return RepRegClassCostForVT[VT.SimpleTy]; 00342 } 00343 00344 /// Return true if the target has native support for the specified value type. 00345 /// This means that it has a register that directly holds it without 00346 /// promotions or expansions. 00347 bool isTypeLegal(EVT VT) const { 00348 assert(!VT.isSimple() || 00349 (unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT)); 00350 return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != nullptr; 00351 } 00352 00353 class ValueTypeActionImpl { 00354 /// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum 00355 /// that indicates how instruction selection should deal with the type. 00356 uint8_t ValueTypeActions[MVT::LAST_VALUETYPE]; 00357 00358 public: 00359 ValueTypeActionImpl() { 00360 std::fill(std::begin(ValueTypeActions), std::end(ValueTypeActions), 0); 00361 } 00362 00363 LegalizeTypeAction getTypeAction(MVT VT) const { 00364 return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy]; 00365 } 00366 00367 void setTypeAction(MVT VT, LegalizeTypeAction Action) { 00368 unsigned I = VT.SimpleTy; 00369 ValueTypeActions[I] = Action; 00370 } 00371 }; 00372 00373 const ValueTypeActionImpl &getValueTypeActions() const { 00374 return ValueTypeActions; 00375 } 00376 00377 /// Return how we should legalize values of this type, either it is already 00378 /// legal (return 'Legal') or we need to promote it to a larger type (return 00379 /// 'Promote'), or we need to expand it into multiple registers of smaller 00380 /// integer type (return 'Expand'). 'Custom' is not an option. 00381 LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const { 00382 return getTypeConversion(Context, VT).first; 00383 } 00384 LegalizeTypeAction getTypeAction(MVT VT) const { 00385 return ValueTypeActions.getTypeAction(VT); 00386 } 00387 00388 /// For types supported by the target, this is an identity function. For 00389 /// types that must be promoted to larger types, this returns the larger type 00390 /// to promote to. For integer types that are larger than the largest integer 00391 /// register, this contains one step in the expansion to get to the smaller 00392 /// register. For illegal floating point types, this returns the integer type 00393 /// to transform to. 00394 EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const { 00395 return getTypeConversion(Context, VT).second; 00396 } 00397 00398 /// For types supported by the target, this is an identity function. For 00399 /// types that must be expanded (i.e. integer types that are larger than the 00400 /// largest integer register or illegal floating point types), this returns 00401 /// the largest legal type it will be expanded to. 00402 EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const { 00403 assert(!VT.isVector()); 00404 while (true) { 00405 switch (getTypeAction(Context, VT)) { 00406 case TypeLegal: 00407 return VT; 00408 case TypeExpandInteger: 00409 VT = getTypeToTransformTo(Context, VT); 00410 break; 00411 default: 00412 llvm_unreachable("Type is not legal nor is it to be expanded!"); 00413 } 00414 } 00415 } 00416 00417 /// Vector types are broken down into some number of legal first class types. 00418 /// For example, EVT::v8f32 maps to 2 EVT::v4f32 with Altivec or SSE1, or 8 00419 /// promoted EVT::f64 values with the X86 FP stack. Similarly, EVT::v2i64 00420 /// turns into 4 EVT::i32 values with both PPC and X86. 00421 /// 00422 /// This method returns the number of registers needed, and the VT for each 00423 /// register. It also returns the VT and quantity of the intermediate values 00424 /// before they are promoted/expanded. 00425 unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 00426 EVT &IntermediateVT, 00427 unsigned &NumIntermediates, 00428 MVT &RegisterVT) const; 00429 00430 struct IntrinsicInfo { 00431 unsigned opc; // target opcode 00432 EVT memVT; // memory VT 00433 const Value* ptrVal; // value representing memory location 00434 int offset; // offset off of ptrVal 00435 unsigned size; // the size of the memory location 00436 // (taken from memVT if zero) 00437 unsigned align; // alignment 00438 bool vol; // is volatile? 00439 bool readMem; // reads memory? 00440 bool writeMem; // writes memory? 00441 00442 IntrinsicInfo() : opc(0), ptrVal(nullptr), offset(0), size(0), align(1), 00443 vol(false), readMem(false), writeMem(false) {} 00444 }; 00445 00446 /// Given an intrinsic, checks if on the target the intrinsic will need to map 00447 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns 00448 /// true and store the intrinsic information into the IntrinsicInfo that was 00449 /// passed to the function. 00450 virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &, 00451 unsigned /*Intrinsic*/) const { 00452 return false; 00453 } 00454 00455 /// Returns true if the target can instruction select the specified FP 00456 /// immediate natively. If false, the legalizer will materialize the FP 00457 /// immediate as a load from a constant pool. 00458 virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const { 00459 return false; 00460 } 00461 00462 /// Targets can use this to indicate that they only support *some* 00463 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a 00464 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to be 00465 /// legal. 00466 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 00467 EVT /*VT*/) const { 00468 return true; 00469 } 00470 00471 /// Returns true if the operation can trap for the value type. 00472 /// 00473 /// VT must be a legal type. By default, we optimistically assume most 00474 /// operations don't trap except for divide and remainder. 00475 virtual bool canOpTrap(unsigned Op, EVT VT) const; 00476 00477 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to 00478 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to replace 00479 /// a VAND with a constant pool entry. 00480 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 00481 EVT /*VT*/) const { 00482 return false; 00483 } 00484 00485 /// Return how this operation should be treated: either it is legal, needs to 00486 /// be promoted to a larger size, needs to be expanded to some other code 00487 /// sequence, or the target has a custom expander for it. 00488 LegalizeAction getOperationAction(unsigned Op, EVT VT) const { 00489 if (VT.isExtended()) return Expand; 00490 // If a target-specific SDNode requires legalization, require the target 00491 // to provide custom legalization for it. 00492 if (Op > array_lengthof(OpActions[0])) return Custom; 00493 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy; 00494 return (LegalizeAction)OpActions[I][Op]; 00495 } 00496 00497 /// Return true if the specified operation is legal on this target or can be 00498 /// made legal with custom lowering. This is used to help guide high-level 00499 /// lowering decisions. 00500 bool isOperationLegalOrCustom(unsigned Op, EVT VT) const { 00501 return (VT == MVT::Other || isTypeLegal(VT)) && 00502 (getOperationAction(Op, VT) == Legal || 00503 getOperationAction(Op, VT) == Custom); 00504 } 00505 00506 /// Return true if the specified operation is legal on this target or can be 00507 /// made legal using promotion. This is used to help guide high-level lowering 00508 /// decisions. 00509 bool isOperationLegalOrPromote(unsigned Op, EVT VT) const { 00510 return (VT == MVT::Other || isTypeLegal(VT)) && 00511 (getOperationAction(Op, VT) == Legal || 00512 getOperationAction(Op, VT) == Promote); 00513 } 00514 00515 /// Return true if the specified operation is illegal on this target or 00516 /// unlikely to be made legal with custom lowering. This is used to help guide 00517 /// high-level lowering decisions. 00518 bool isOperationExpand(unsigned Op, EVT VT) const { 00519 return (!isTypeLegal(VT) || getOperationAction(Op, VT) == Expand); 00520 } 00521 00522 /// Return true if the specified operation is legal on this target. 00523 bool isOperationLegal(unsigned Op, EVT VT) const { 00524 return (VT == MVT::Other || isTypeLegal(VT)) && 00525 getOperationAction(Op, VT) == Legal; 00526 } 00527 00528 /// Return how this load with extension should be treated: either it is legal, 00529 /// needs to be promoted to a larger size, needs to be expanded to some other 00530 /// code sequence, or the target has a custom expander for it. 00531 LegalizeAction getLoadExtAction(unsigned ExtType, EVT VT) const { 00532 if (VT.isExtended()) return Expand; 00533 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy; 00534 assert(ExtType < ISD::LAST_LOADEXT_TYPE && I < MVT::LAST_VALUETYPE && 00535 "Table isn't big enough!"); 00536 return (LegalizeAction)LoadExtActions[I][ExtType]; 00537 } 00538 00539 /// Return true if the specified load with extension is legal on this target. 00540 bool isLoadExtLegal(unsigned ExtType, EVT VT) const { 00541 return VT.isSimple() && 00542 getLoadExtAction(ExtType, VT.getSimpleVT()) == Legal; 00543 } 00544 00545 /// Return how this store with truncation should be treated: either it is 00546 /// legal, needs to be promoted to a larger size, needs to be expanded to some 00547 /// other code sequence, or the target has a custom expander for it. 00548 LegalizeAction getTruncStoreAction(EVT ValVT, EVT MemVT) const { 00549 if (ValVT.isExtended() || MemVT.isExtended()) return Expand; 00550 unsigned ValI = (unsigned) ValVT.getSimpleVT().SimpleTy; 00551 unsigned MemI = (unsigned) MemVT.getSimpleVT().SimpleTy; 00552 assert(ValI < MVT::LAST_VALUETYPE && MemI < MVT::LAST_VALUETYPE && 00553 "Table isn't big enough!"); 00554 return (LegalizeAction)TruncStoreActions[ValI][MemI]; 00555 } 00556 00557 /// Return true if the specified store with truncation is legal on this 00558 /// target. 00559 bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const { 00560 return isTypeLegal(ValVT) && MemVT.isSimple() && 00561 getTruncStoreAction(ValVT.getSimpleVT(), MemVT.getSimpleVT()) == Legal; 00562 } 00563 00564 /// Return how the indexed load should be treated: either it is legal, needs 00565 /// to be promoted to a larger size, needs to be expanded to some other code 00566 /// sequence, or the target has a custom expander for it. 00567 LegalizeAction 00568 getIndexedLoadAction(unsigned IdxMode, MVT VT) const { 00569 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE && 00570 "Table isn't big enough!"); 00571 unsigned Ty = (unsigned)VT.SimpleTy; 00572 return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4); 00573 } 00574 00575 /// Return true if the specified indexed load is legal on this target. 00576 bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const { 00577 return VT.isSimple() && 00578 (getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Legal || 00579 getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Custom); 00580 } 00581 00582 /// Return how the indexed store should be treated: either it is legal, needs 00583 /// to be promoted to a larger size, needs to be expanded to some other code 00584 /// sequence, or the target has a custom expander for it. 00585 LegalizeAction 00586 getIndexedStoreAction(unsigned IdxMode, MVT VT) const { 00587 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE && 00588 "Table isn't big enough!"); 00589 unsigned Ty = (unsigned)VT.SimpleTy; 00590 return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f); 00591 } 00592 00593 /// Return true if the specified indexed load is legal on this target. 00594 bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const { 00595 return VT.isSimple() && 00596 (getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Legal || 00597 getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Custom); 00598 } 00599 00600 /// Return how the condition code should be treated: either it is legal, needs 00601 /// to be expanded to some other code sequence, or the target has a custom 00602 /// expander for it. 00603 LegalizeAction 00604 getCondCodeAction(ISD::CondCode CC, MVT VT) const { 00605 assert((unsigned)CC < array_lengthof(CondCodeActions) && 00606 ((unsigned)VT.SimpleTy >> 4) < array_lengthof(CondCodeActions[0]) && 00607 "Table isn't big enough!"); 00608 // See setCondCodeAction for how this is encoded. 00609 uint32_t Shift = 2 * (VT.SimpleTy & 0xF); 00610 uint32_t Value = CondCodeActions[CC][VT.SimpleTy >> 4]; 00611 LegalizeAction Action = (LegalizeAction) ((Value >> Shift) & 0x3); 00612 assert(Action != Promote && "Can't promote condition code!"); 00613 return Action; 00614 } 00615 00616 /// Return true if the specified condition code is legal on this target. 00617 bool isCondCodeLegal(ISD::CondCode CC, MVT VT) const { 00618 return 00619 getCondCodeAction(CC, VT) == Legal || 00620 getCondCodeAction(CC, VT) == Custom; 00621 } 00622 00623 00624 /// If the action for this operation is to promote, this method returns the 00625 /// ValueType to promote to. 00626 MVT getTypeToPromoteTo(unsigned Op, MVT VT) const { 00627 assert(getOperationAction(Op, VT) == Promote && 00628 "This operation isn't promoted!"); 00629 00630 // See if this has an explicit type specified. 00631 std::map<std::pair<unsigned, MVT::SimpleValueType>, 00632 MVT::SimpleValueType>::const_iterator PTTI = 00633 PromoteToType.find(std::make_pair(Op, VT.SimpleTy)); 00634 if (PTTI != PromoteToType.end()) return PTTI->second; 00635 00636 assert((VT.isInteger() || VT.isFloatingPoint()) && 00637 "Cannot autopromote this type, add it with AddPromotedToType."); 00638 00639 MVT NVT = VT; 00640 do { 00641 NVT = (MVT::SimpleValueType)(NVT.SimpleTy+1); 00642 assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid && 00643 "Didn't find type to promote to!"); 00644 } while (!isTypeLegal(NVT) || 00645 getOperationAction(Op, NVT) == Promote); 00646 return NVT; 00647 } 00648 00649 /// Return the EVT corresponding to this LLVM type. This is fixed by the LLVM 00650 /// operations except for the pointer size. If AllowUnknown is true, this 00651 /// will return MVT::Other for types with no EVT counterpart (e.g. structs), 00652 /// otherwise it will assert. 00653 EVT getValueType(Type *Ty, bool AllowUnknown = false) const { 00654 // Lower scalar pointers to native pointer types. 00655 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 00656 return getPointerTy(PTy->getAddressSpace()); 00657 00658 if (Ty->isVectorTy()) { 00659 VectorType *VTy = cast<VectorType>(Ty); 00660 Type *Elm = VTy->getElementType(); 00661 // Lower vectors of pointers to native pointer types. 00662 if (PointerType *PT = dyn_cast<PointerType>(Elm)) { 00663 EVT PointerTy(getPointerTy(PT->getAddressSpace())); 00664 Elm = PointerTy.getTypeForEVT(Ty->getContext()); 00665 } 00666 00667 return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false), 00668 VTy->getNumElements()); 00669 } 00670 return EVT::getEVT(Ty, AllowUnknown); 00671 } 00672 00673 /// Return the MVT corresponding to this LLVM type. See getValueType. 00674 MVT getSimpleValueType(Type *Ty, bool AllowUnknown = false) const { 00675 return getValueType(Ty, AllowUnknown).getSimpleVT(); 00676 } 00677 00678 /// Return the desired alignment for ByVal or InAlloca aggregate function 00679 /// arguments in the caller parameter area. This is the actual alignment, not 00680 /// its logarithm. 00681 virtual unsigned getByValTypeAlignment(Type *Ty) const; 00682 00683 /// Return the type of registers that this ValueType will eventually require. 00684 MVT getRegisterType(MVT VT) const { 00685 assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT)); 00686 return RegisterTypeForVT[VT.SimpleTy]; 00687 } 00688 00689 /// Return the type of registers that this ValueType will eventually require. 00690 MVT getRegisterType(LLVMContext &Context, EVT VT) const { 00691 if (VT.isSimple()) { 00692 assert((unsigned)VT.getSimpleVT().SimpleTy < 00693 array_lengthof(RegisterTypeForVT)); 00694 return RegisterTypeForVT[VT.getSimpleVT().SimpleTy]; 00695 } 00696 if (VT.isVector()) { 00697 EVT VT1; 00698 MVT RegisterVT; 00699 unsigned NumIntermediates; 00700 (void)getVectorTypeBreakdown(Context, VT, VT1, 00701 NumIntermediates, RegisterVT); 00702 return RegisterVT; 00703 } 00704 if (VT.isInteger()) { 00705 return getRegisterType(Context, getTypeToTransformTo(Context, VT)); 00706 } 00707 llvm_unreachable("Unsupported extended type!"); 00708 } 00709 00710 /// Return the number of registers that this ValueType will eventually 00711 /// require. 00712 /// 00713 /// This is one for any types promoted to live in larger registers, but may be 00714 /// more than one for types (like i64) that are split into pieces. For types 00715 /// like i140, which are first promoted then expanded, it is the number of 00716 /// registers needed to hold all the bits of the original type. For an i140 00717 /// on a 32 bit machine this means 5 registers. 00718 unsigned getNumRegisters(LLVMContext &Context, EVT VT) const { 00719 if (VT.isSimple()) { 00720 assert((unsigned)VT.getSimpleVT().SimpleTy < 00721 array_lengthof(NumRegistersForVT)); 00722 return NumRegistersForVT[VT.getSimpleVT().SimpleTy]; 00723 } 00724 if (VT.isVector()) { 00725 EVT VT1; 00726 MVT VT2; 00727 unsigned NumIntermediates; 00728 return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2); 00729 } 00730 if (VT.isInteger()) { 00731 unsigned BitWidth = VT.getSizeInBits(); 00732 unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits(); 00733 return (BitWidth + RegWidth - 1) / RegWidth; 00734 } 00735 llvm_unreachable("Unsupported extended type!"); 00736 } 00737 00738 /// If true, then instruction selection should seek to shrink the FP constant 00739 /// of the specified type to a smaller type in order to save space and / or 00740 /// reduce runtime. 00741 virtual bool ShouldShrinkFPConstant(EVT) const { return true; } 00742 00743 /// When splitting a value of the specified type into parts, does the Lo 00744 /// or Hi part come first? This usually follows the endianness, except 00745 /// for ppcf128, where the Hi part always comes first. 00746 bool hasBigEndianPartOrdering(EVT VT) const { 00747 return isBigEndian() || VT == MVT::ppcf128; 00748 } 00749 00750 /// If true, the target has custom DAG combine transformations that it can 00751 /// perform for the specified node. 00752 bool hasTargetDAGCombine(ISD::NodeType NT) const { 00753 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 00754 return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7)); 00755 } 00756 00757 /// \brief Get maximum # of store operations permitted for llvm.memset 00758 /// 00759 /// This function returns the maximum number of store operations permitted 00760 /// to replace a call to llvm.memset. The value is set by the target at the 00761 /// performance threshold for such a replacement. If OptSize is true, 00762 /// return the limit for functions that have OptSize attribute. 00763 unsigned getMaxStoresPerMemset(bool OptSize) const { 00764 return OptSize ? MaxStoresPerMemsetOptSize : MaxStoresPerMemset; 00765 } 00766 00767 /// \brief Get maximum # of store operations permitted for llvm.memcpy 00768 /// 00769 /// This function returns the maximum number of store operations permitted 00770 /// to replace a call to llvm.memcpy. The value is set by the target at the 00771 /// performance threshold for such a replacement. If OptSize is true, 00772 /// return the limit for functions that have OptSize attribute. 00773 unsigned getMaxStoresPerMemcpy(bool OptSize) const { 00774 return OptSize ? MaxStoresPerMemcpyOptSize : MaxStoresPerMemcpy; 00775 } 00776 00777 /// \brief Get maximum # of store operations permitted for llvm.memmove 00778 /// 00779 /// This function returns the maximum number of store operations permitted 00780 /// to replace a call to llvm.memmove. The value is set by the target at the 00781 /// performance threshold for such a replacement. If OptSize is true, 00782 /// return the limit for functions that have OptSize attribute. 00783 unsigned getMaxStoresPerMemmove(bool OptSize) const { 00784 return OptSize ? MaxStoresPerMemmoveOptSize : MaxStoresPerMemmove; 00785 } 00786 00787 /// \brief Determine if the target supports unaligned memory accesses. 00788 /// 00789 /// This function returns true if the target allows unaligned memory accesses 00790 /// of the specified type in the given address space. If true, it also returns 00791 /// whether the unaligned memory access is "fast" in the last argument by 00792 /// reference. This is used, for example, in situations where an array 00793 /// copy/move/set is converted to a sequence of store operations. Its use 00794 /// helps to ensure that such replacements don't generate code that causes an 00795 /// alignment error (trap) on the target machine. 00796 virtual bool allowsMisalignedMemoryAccesses(EVT, 00797 unsigned AddrSpace = 0, 00798 unsigned Align = 1, 00799 bool * /*Fast*/ = nullptr) const { 00800 return false; 00801 } 00802 00803 /// Returns the target specific optimal type for load and store operations as 00804 /// a result of memset, memcpy, and memmove lowering. 00805 /// 00806 /// If DstAlign is zero that means it's safe to destination alignment can 00807 /// satisfy any constraint. Similarly if SrcAlign is zero it means there isn't 00808 /// a need to check it against alignment requirement, probably because the 00809 /// source does not need to be loaded. If 'IsMemset' is true, that means it's 00810 /// expanding a memset. If 'ZeroMemset' is true, that means it's a memset of 00811 /// zero. 'MemcpyStrSrc' indicates whether the memcpy source is constant so it 00812 /// does not need to be loaded. It returns EVT::Other if the type should be 00813 /// determined using generic target-independent logic. 00814 virtual EVT getOptimalMemOpType(uint64_t /*Size*/, 00815 unsigned /*DstAlign*/, unsigned /*SrcAlign*/, 00816 bool /*IsMemset*/, 00817 bool /*ZeroMemset*/, 00818 bool /*MemcpyStrSrc*/, 00819 MachineFunction &/*MF*/) const { 00820 return MVT::Other; 00821 } 00822 00823 /// Returns true if it's safe to use load / store of the specified type to 00824 /// expand memcpy / memset inline. 00825 /// 00826 /// This is mostly true for all types except for some special cases. For 00827 /// example, on X86 targets without SSE2 f64 load / store are done with fldl / 00828 /// fstpl which also does type conversion. Note the specified type doesn't 00829 /// have to be legal as the hook is used before type legalization. 00830 virtual bool isSafeMemOpType(MVT /*VT*/) const { return true; } 00831 00832 /// Determine if we should use _setjmp or setjmp to implement llvm.setjmp. 00833 bool usesUnderscoreSetJmp() const { 00834 return UseUnderscoreSetJmp; 00835 } 00836 00837 /// Determine if we should use _longjmp or longjmp to implement llvm.longjmp. 00838 bool usesUnderscoreLongJmp() const { 00839 return UseUnderscoreLongJmp; 00840 } 00841 00842 /// Return integer threshold on number of blocks to use jump tables rather 00843 /// than if sequence. 00844 int getMinimumJumpTableEntries() const { 00845 return MinimumJumpTableEntries; 00846 } 00847 00848 /// If a physical register, this specifies the register that 00849 /// llvm.savestack/llvm.restorestack should save and restore. 00850 unsigned getStackPointerRegisterToSaveRestore() const { 00851 return StackPointerRegisterToSaveRestore; 00852 } 00853 00854 /// If a physical register, this returns the register that receives the 00855 /// exception address on entry to a landing pad. 00856 unsigned getExceptionPointerRegister() const { 00857 return ExceptionPointerRegister; 00858 } 00859 00860 /// If a physical register, this returns the register that receives the 00861 /// exception typeid on entry to a landing pad. 00862 unsigned getExceptionSelectorRegister() const { 00863 return ExceptionSelectorRegister; 00864 } 00865 00866 /// Returns the target's jmp_buf size in bytes (if never set, the default is 00867 /// 200) 00868 unsigned getJumpBufSize() const { 00869 return JumpBufSize; 00870 } 00871 00872 /// Returns the target's jmp_buf alignment in bytes (if never set, the default 00873 /// is 0) 00874 unsigned getJumpBufAlignment() const { 00875 return JumpBufAlignment; 00876 } 00877 00878 /// Return the minimum stack alignment of an argument. 00879 unsigned getMinStackArgumentAlignment() const { 00880 return MinStackArgumentAlignment; 00881 } 00882 00883 /// Return the minimum function alignment. 00884 unsigned getMinFunctionAlignment() const { 00885 return MinFunctionAlignment; 00886 } 00887 00888 /// Return the preferred function alignment. 00889 unsigned getPrefFunctionAlignment() const { 00890 return PrefFunctionAlignment; 00891 } 00892 00893 /// Return the preferred loop alignment. 00894 unsigned getPrefLoopAlignment() const { 00895 return PrefLoopAlignment; 00896 } 00897 00898 /// Return whether the DAG builder should automatically insert fences and 00899 /// reduce ordering for atomics. 00900 bool getInsertFencesForAtomic() const { 00901 return InsertFencesForAtomic; 00902 } 00903 00904 /// Return true if the target stores stack protector cookies at a fixed offset 00905 /// in some non-standard address space, and populates the address space and 00906 /// offset as appropriate. 00907 virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/, 00908 unsigned &/*Offset*/) const { 00909 return false; 00910 } 00911 00912 /// Returns the maximal possible offset which can be used for loads / stores 00913 /// from the global. 00914 virtual unsigned getMaximalGlobalOffset() const { 00915 return 0; 00916 } 00917 00918 /// Returns true if a cast between SrcAS and DestAS is a noop. 00919 virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const { 00920 return false; 00921 } 00922 00923 //===--------------------------------------------------------------------===// 00924 /// \name Helpers for TargetTransformInfo implementations 00925 /// @{ 00926 00927 /// Get the ISD node that corresponds to the Instruction class opcode. 00928 int InstructionOpcodeToISD(unsigned Opcode) const; 00929 00930 /// Estimate the cost of type-legalization and the legalized type. 00931 std::pair<unsigned, MVT> getTypeLegalizationCost(Type *Ty) const; 00932 00933 /// @} 00934 00935 //===--------------------------------------------------------------------===// 00936 /// \name Helpers for atomic expansion. 00937 /// @{ 00938 00939 /// True if AtomicExpandPass should use emitLoadLinked/emitStoreConditional 00940 /// and expand AtomicCmpXchgInst. 00941 virtual bool hasLoadLinkedStoreConditional() const { return false; } 00942 00943 /// Perform a load-linked operation on Addr, returning a "Value *" with the 00944 /// corresponding pointee type. This may entail some non-trivial operations to 00945 /// truncate or reconstruct types that will be illegal in the backend. See 00946 /// ARMISelLowering for an example implementation. 00947 virtual Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr, 00948 AtomicOrdering Ord) const { 00949 llvm_unreachable("Load linked unimplemented on this target"); 00950 } 00951 00952 /// Perform a store-conditional operation to Addr. Return the status of the 00953 /// store. This should be 0 if the store succeeded, non-zero otherwise. 00954 virtual Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val, 00955 Value *Addr, AtomicOrdering Ord) const { 00956 llvm_unreachable("Store conditional unimplemented on this target"); 00957 } 00958 00959 /// Inserts in the IR a target-specific intrinsic specifying a fence. 00960 /// It is called by AtomicExpandPass before expanding an 00961 /// AtomicRMW/AtomicCmpXchg/AtomicStore/AtomicLoad. 00962 /// RMW and CmpXchg set both IsStore and IsLoad to true. 00963 /// Backends with !getInsertFencesForAtomic() should keep a no-op here. 00964 virtual void emitLeadingFence(IRBuilder<> &Builder, AtomicOrdering Ord, 00965 bool IsStore, bool IsLoad) const { 00966 assert(!getInsertFencesForAtomic()); 00967 } 00968 00969 /// Inserts in the IR a target-specific intrinsic specifying a fence. 00970 /// It is called by AtomicExpandPass after expanding an 00971 /// AtomicRMW/AtomicCmpXchg/AtomicStore/AtomicLoad. 00972 /// RMW and CmpXchg set both IsStore and IsLoad to true. 00973 /// Backends with !getInsertFencesForAtomic() should keep a no-op here. 00974 virtual void emitTrailingFence(IRBuilder<> &Builder, AtomicOrdering Ord, 00975 bool IsStore, bool IsLoad) const { 00976 assert(!getInsertFencesForAtomic()); 00977 } 00978 00979 /// Returns true if the given (atomic) store should be expanded by the 00980 /// IR-level AtomicExpand pass into an "atomic xchg" which ignores its input. 00981 virtual bool shouldExpandAtomicStoreInIR(StoreInst *SI) const { 00982 return false; 00983 } 00984 00985 /// Returns true if the given (atomic) load should be expanded by the 00986 /// IR-level AtomicExpand pass into a load-linked instruction 00987 /// (through emitLoadLinked()). 00988 virtual bool shouldExpandAtomicLoadInIR(LoadInst *LI) const { return false; } 00989 00990 /// Returns true if the given AtomicRMW should be expanded by the 00991 /// IR-level AtomicExpand pass into a loop using LoadLinked/StoreConditional. 00992 virtual bool shouldExpandAtomicRMWInIR(AtomicRMWInst *RMWI) const { 00993 return false; 00994 } 00995 00996 //===--------------------------------------------------------------------===// 00997 // TargetLowering Configuration Methods - These methods should be invoked by 00998 // the derived class constructor to configure this object for the target. 00999 // 01000 01001 /// \brief Reset the operation actions based on target options. 01002 virtual void resetOperationActions() {} 01003 01004 protected: 01005 /// Specify how the target extends the result of integer and floating point 01006 /// boolean values from i1 to a wider type. See getBooleanContents. 01007 void setBooleanContents(BooleanContent Ty) { 01008 BooleanContents = Ty; 01009 BooleanFloatContents = Ty; 01010 } 01011 01012 /// Specify how the target extends the result of integer and floating point 01013 /// boolean values from i1 to a wider type. See getBooleanContents. 01014 void setBooleanContents(BooleanContent IntTy, BooleanContent FloatTy) { 01015 BooleanContents = IntTy; 01016 BooleanFloatContents = FloatTy; 01017 } 01018 01019 /// Specify how the target extends the result of a vector boolean value from a 01020 /// vector of i1 to a wider type. See getBooleanContents. 01021 void setBooleanVectorContents(BooleanContent Ty) { 01022 BooleanVectorContents = Ty; 01023 } 01024 01025 /// Specify the target scheduling preference. 01026 void setSchedulingPreference(Sched::Preference Pref) { 01027 SchedPreferenceInfo = Pref; 01028 } 01029 01030 /// Indicate whether this target prefers to use _setjmp to implement 01031 /// llvm.setjmp or the version without _. Defaults to false. 01032 void setUseUnderscoreSetJmp(bool Val) { 01033 UseUnderscoreSetJmp = Val; 01034 } 01035 01036 /// Indicate whether this target prefers to use _longjmp to implement 01037 /// llvm.longjmp or the version without _. Defaults to false. 01038 void setUseUnderscoreLongJmp(bool Val) { 01039 UseUnderscoreLongJmp = Val; 01040 } 01041 01042 /// Indicate the number of blocks to generate jump tables rather than if 01043 /// sequence. 01044 void setMinimumJumpTableEntries(int Val) { 01045 MinimumJumpTableEntries = Val; 01046 } 01047 01048 /// If set to a physical register, this specifies the register that 01049 /// llvm.savestack/llvm.restorestack should save and restore. 01050 void setStackPointerRegisterToSaveRestore(unsigned R) { 01051 StackPointerRegisterToSaveRestore = R; 01052 } 01053 01054 /// If set to a physical register, this sets the register that receives the 01055 /// exception address on entry to a landing pad. 01056 void setExceptionPointerRegister(unsigned R) { 01057 ExceptionPointerRegister = R; 01058 } 01059 01060 /// If set to a physical register, this sets the register that receives the 01061 /// exception typeid on entry to a landing pad. 01062 void setExceptionSelectorRegister(unsigned R) { 01063 ExceptionSelectorRegister = R; 01064 } 01065 01066 /// Tells the code generator not to expand operations into sequences that use 01067 /// the select operations if possible. 01068 void setSelectIsExpensive(bool isExpensive = true) { 01069 SelectIsExpensive = isExpensive; 01070 } 01071 01072 /// Tells the code generator that the target has multiple (allocatable) 01073 /// condition registers that can be used to store the results of comparisons 01074 /// for use by selects and conditional branches. With multiple condition 01075 /// registers, the code generator will not aggressively sink comparisons into 01076 /// the blocks of their users. 01077 void setHasMultipleConditionRegisters(bool hasManyRegs = true) { 01078 HasMultipleConditionRegisters = hasManyRegs; 01079 } 01080 01081 /// Tells the code generator that the target has BitExtract instructions. 01082 /// The code generator will aggressively sink "shift"s into the blocks of 01083 /// their users if the users will generate "and" instructions which can be 01084 /// combined with "shift" to BitExtract instructions. 01085 void setHasExtractBitsInsn(bool hasExtractInsn = true) { 01086 HasExtractBitsInsn = hasExtractInsn; 01087 } 01088 01089 /// Tells the code generator not to expand sequence of operations into a 01090 /// separate sequences that increases the amount of flow control. 01091 void setJumpIsExpensive(bool isExpensive = true) { 01092 JumpIsExpensive = isExpensive; 01093 } 01094 01095 /// Tells the code generator that integer divide is expensive, and if 01096 /// possible, should be replaced by an alternate sequence of instructions not 01097 /// containing an integer divide. 01098 void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; } 01099 01100 /// Tells the code generator that this target supports floating point 01101 /// exceptions and cares about preserving floating point exception behavior. 01102 void setHasFloatingPointExceptions(bool FPExceptions = true) { 01103 HasFloatingPointExceptions = FPExceptions; 01104 } 01105 01106 /// Tells the code generator which bitwidths to bypass. 01107 void addBypassSlowDiv(unsigned int SlowBitWidth, unsigned int FastBitWidth) { 01108 BypassSlowDivWidths[SlowBitWidth] = FastBitWidth; 01109 } 01110 01111 /// Tells the code generator that it shouldn't generate sra/srl/add/sra for a 01112 /// signed divide by power of two; let the target handle it. 01113 void setPow2SDivIsCheap(bool isCheap = true) { Pow2SDivIsCheap = isCheap; } 01114 01115 /// Add the specified register class as an available regclass for the 01116 /// specified value type. This indicates the selector can handle values of 01117 /// that class natively. 01118 void addRegisterClass(MVT VT, const TargetRegisterClass *RC) { 01119 assert((unsigned)VT.SimpleTy < array_lengthof(RegClassForVT)); 01120 AvailableRegClasses.push_back(std::make_pair(VT, RC)); 01121 RegClassForVT[VT.SimpleTy] = RC; 01122 } 01123 01124 /// Remove all register classes. 01125 void clearRegisterClasses() { 01126 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE * sizeof(TargetRegisterClass*)); 01127 01128 AvailableRegClasses.clear(); 01129 } 01130 01131 /// \brief Remove all operation actions. 01132 void clearOperationActions() { 01133 } 01134 01135 /// Return the largest legal super-reg register class of the register class 01136 /// for the specified type and its associated "cost". 01137 virtual std::pair<const TargetRegisterClass*, uint8_t> 01138 findRepresentativeClass(MVT VT) const; 01139 01140 /// Once all of the register classes are added, this allows us to compute 01141 /// derived properties we expose. 01142 void computeRegisterProperties(); 01143 01144 /// Indicate that the specified operation does not work with the specified 01145 /// type and indicate what to do about it. 01146 void setOperationAction(unsigned Op, MVT VT, 01147 LegalizeAction Action) { 01148 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!"); 01149 OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action; 01150 } 01151 01152 /// Indicate that the specified load with extension does not work with the 01153 /// specified type and indicate what to do about it. 01154 void setLoadExtAction(unsigned ExtType, MVT VT, 01155 LegalizeAction Action) { 01156 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE && 01157 "Table isn't big enough!"); 01158 LoadExtActions[VT.SimpleTy][ExtType] = (uint8_t)Action; 01159 } 01160 01161 /// Indicate that the specified truncating store does not work with the 01162 /// specified type and indicate what to do about it. 01163 void setTruncStoreAction(MVT ValVT, MVT MemVT, 01164 LegalizeAction Action) { 01165 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE && 01166 "Table isn't big enough!"); 01167 TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action; 01168 } 01169 01170 /// Indicate that the specified indexed load does or does not work with the 01171 /// specified type and indicate what to do abort it. 01172 /// 01173 /// NOTE: All indexed mode loads are initialized to Expand in 01174 /// TargetLowering.cpp 01175 void setIndexedLoadAction(unsigned IdxMode, MVT VT, 01176 LegalizeAction Action) { 01177 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE && 01178 (unsigned)Action < 0xf && "Table isn't big enough!"); 01179 // Load action are kept in the upper half. 01180 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0; 01181 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4; 01182 } 01183 01184 /// Indicate that the specified indexed store does or does not work with the 01185 /// specified type and indicate what to do about it. 01186 /// 01187 /// NOTE: All indexed mode stores are initialized to Expand in 01188 /// TargetLowering.cpp 01189 void setIndexedStoreAction(unsigned IdxMode, MVT VT, 01190 LegalizeAction Action) { 01191 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE && 01192 (unsigned)Action < 0xf && "Table isn't big enough!"); 01193 // Store action are kept in the lower half. 01194 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f; 01195 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action); 01196 } 01197 01198 /// Indicate that the specified condition code is or isn't supported on the 01199 /// target and indicate what to do about it. 01200 void setCondCodeAction(ISD::CondCode CC, MVT VT, 01201 LegalizeAction Action) { 01202 assert(VT < MVT::LAST_VALUETYPE && 01203 (unsigned)CC < array_lengthof(CondCodeActions) && 01204 "Table isn't big enough!"); 01205 /// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 32-bit 01206 /// value and the upper 27 bits index into the second dimension of the array 01207 /// to select what 32-bit value to use. 01208 uint32_t Shift = 2 * (VT.SimpleTy & 0xF); 01209 CondCodeActions[CC][VT.SimpleTy >> 4] &= ~((uint32_t)0x3 << Shift); 01210 CondCodeActions[CC][VT.SimpleTy >> 4] |= (uint32_t)Action << Shift; 01211 } 01212 01213 /// If Opc/OrigVT is specified as being promoted, the promotion code defaults 01214 /// to trying a larger integer/fp until it can find one that works. If that 01215 /// default is insufficient, this method can be used by the target to override 01216 /// the default. 01217 void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) { 01218 PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy; 01219 } 01220 01221 /// Targets should invoke this method for each target independent node that 01222 /// they want to provide a custom DAG combiner for by implementing the 01223 /// PerformDAGCombine virtual method. 01224 void setTargetDAGCombine(ISD::NodeType NT) { 01225 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 01226 TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7); 01227 } 01228 01229 /// Set the target's required jmp_buf buffer size (in bytes); default is 200 01230 void setJumpBufSize(unsigned Size) { 01231 JumpBufSize = Size; 01232 } 01233 01234 /// Set the target's required jmp_buf buffer alignment (in bytes); default is 01235 /// 0 01236 void setJumpBufAlignment(unsigned Align) { 01237 JumpBufAlignment = Align; 01238 } 01239 01240 /// Set the target's minimum function alignment (in log2(bytes)) 01241 void setMinFunctionAlignment(unsigned Align) { 01242 MinFunctionAlignment = Align; 01243 } 01244 01245 /// Set the target's preferred function alignment. This should be set if 01246 /// there is a performance benefit to higher-than-minimum alignment (in 01247 /// log2(bytes)) 01248 void setPrefFunctionAlignment(unsigned Align) { 01249 PrefFunctionAlignment = Align; 01250 } 01251 01252 /// Set the target's preferred loop alignment. Default alignment is zero, it 01253 /// means the target does not care about loop alignment. The alignment is 01254 /// specified in log2(bytes). 01255 void setPrefLoopAlignment(unsigned Align) { 01256 PrefLoopAlignment = Align; 01257 } 01258 01259 /// Set the minimum stack alignment of an argument (in log2(bytes)). 01260 void setMinStackArgumentAlignment(unsigned Align) { 01261 MinStackArgumentAlignment = Align; 01262 } 01263 01264 /// Set if the DAG builder should automatically insert fences and reduce the 01265 /// order of atomic memory operations to Monotonic. 01266 void setInsertFencesForAtomic(bool fence) { 01267 InsertFencesForAtomic = fence; 01268 } 01269 01270 public: 01271 //===--------------------------------------------------------------------===// 01272 // Addressing mode description hooks (used by LSR etc). 01273 // 01274 01275 /// CodeGenPrepare sinks address calculations into the same BB as Load/Store 01276 /// instructions reading the address. This allows as much computation as 01277 /// possible to be done in the address mode for that operand. This hook lets 01278 /// targets also pass back when this should be done on intrinsics which 01279 /// load/store. 01280 virtual bool GetAddrModeArguments(IntrinsicInst * /*I*/, 01281 SmallVectorImpl<Value*> &/*Ops*/, 01282 Type *&/*AccessTy*/) const { 01283 return false; 01284 } 01285 01286 /// This represents an addressing mode of: 01287 /// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg 01288 /// If BaseGV is null, there is no BaseGV. 01289 /// If BaseOffs is zero, there is no base offset. 01290 /// If HasBaseReg is false, there is no base register. 01291 /// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with 01292 /// no scale. 01293 struct AddrMode { 01294 GlobalValue *BaseGV; 01295 int64_t BaseOffs; 01296 bool HasBaseReg; 01297 int64_t Scale; 01298 AddrMode() : BaseGV(nullptr), BaseOffs(0), HasBaseReg(false), Scale(0) {} 01299 }; 01300 01301 /// Return true if the addressing mode represented by AM is legal for this 01302 /// target, for a load/store of the specified type. 01303 /// 01304 /// The type may be VoidTy, in which case only return true if the addressing 01305 /// mode is legal for a load/store of any legal type. TODO: Handle 01306 /// pre/postinc as well. 01307 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const; 01308 01309 /// \brief Return the cost of the scaling factor used in the addressing mode 01310 /// represented by AM for this target, for a load/store of the specified type. 01311 /// 01312 /// If the AM is supported, the return value must be >= 0. 01313 /// If the AM is not supported, it returns a negative value. 01314 /// TODO: Handle pre/postinc as well. 01315 virtual int getScalingFactorCost(const AddrMode &AM, Type *Ty) const { 01316 // Default: assume that any scaling factor used in a legal AM is free. 01317 if (isLegalAddressingMode(AM, Ty)) return 0; 01318 return -1; 01319 } 01320 01321 /// Return true if the specified immediate is legal icmp immediate, that is 01322 /// the target has icmp instructions which can compare a register against the 01323 /// immediate without having to materialize the immediate into a register. 01324 virtual bool isLegalICmpImmediate(int64_t) const { 01325 return true; 01326 } 01327 01328 /// Return true if the specified immediate is legal add immediate, that is the 01329 /// target has add instructions which can add a register with the immediate 01330 /// without having to materialize the immediate into a register. 01331 virtual bool isLegalAddImmediate(int64_t) const { 01332 return true; 01333 } 01334 01335 /// Return true if it's significantly cheaper to shift a vector by a uniform 01336 /// scalar than by an amount which will vary across each lane. On x86, for 01337 /// example, there is a "psllw" instruction for the former case, but no simple 01338 /// instruction for a general "a << b" operation on vectors. 01339 virtual bool isVectorShiftByScalarCheap(Type *Ty) const { 01340 return false; 01341 } 01342 01343 /// Return true if it's free to truncate a value of type Ty1 to type 01344 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16 01345 /// by referencing its sub-register AX. 01346 virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 01347 return false; 01348 } 01349 01350 /// Return true if a truncation from Ty1 to Ty2 is permitted when deciding 01351 /// whether a call is in tail position. Typically this means that both results 01352 /// would be assigned to the same register or stack slot, but it could mean 01353 /// the target performs adequate checks of its own before proceeding with the 01354 /// tail call. 01355 virtual bool allowTruncateForTailCall(Type * /*Ty1*/, Type * /*Ty2*/) const { 01356 return false; 01357 } 01358 01359 virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const { 01360 return false; 01361 } 01362 01363 /// Return true if any actual instruction that defines a value of type Ty1 01364 /// implicitly zero-extends the value to Ty2 in the result register. 01365 /// 01366 /// This does not necessarily include registers defined in unknown ways, such 01367 /// as incoming arguments, or copies from unknown virtual registers. Also, if 01368 /// isTruncateFree(Ty2, Ty1) is true, this does not necessarily apply to 01369 /// truncate instructions. e.g. on x86-64, all instructions that define 32-bit 01370 /// values implicit zero-extend the result out to 64 bits. 01371 virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 01372 return false; 01373 } 01374 01375 virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const { 01376 return false; 01377 } 01378 01379 /// Return true if the target supplies and combines to a paired load 01380 /// two loaded values of type LoadedType next to each other in memory. 01381 /// RequiredAlignment gives the minimal alignment constraints that must be met 01382 /// to be able to select this paired load. 01383 /// 01384 /// This information is *not* used to generate actual paired loads, but it is 01385 /// used to generate a sequence of loads that is easier to combine into a 01386 /// paired load. 01387 /// For instance, something like this: 01388 /// a = load i64* addr 01389 /// b = trunc i64 a to i32 01390 /// c = lshr i64 a, 32 01391 /// d = trunc i64 c to i32 01392 /// will be optimized into: 01393 /// b = load i32* addr1 01394 /// d = load i32* addr2 01395 /// Where addr1 = addr2 +/- sizeof(i32). 01396 /// 01397 /// In other words, unless the target performs a post-isel load combining, 01398 /// this information should not be provided because it will generate more 01399 /// loads. 01400 virtual bool hasPairedLoad(Type * /*LoadedType*/, 01401 unsigned & /*RequiredAligment*/) const { 01402 return false; 01403 } 01404 01405 virtual bool hasPairedLoad(EVT /*LoadedType*/, 01406 unsigned & /*RequiredAligment*/) const { 01407 return false; 01408 } 01409 01410 /// Return true if zero-extending the specific node Val to type VT2 is free 01411 /// (either because it's implicitly zero-extended such as ARM ldrb / ldrh or 01412 /// because it's folded such as X86 zero-extending loads). 01413 virtual bool isZExtFree(SDValue Val, EVT VT2) const { 01414 return isZExtFree(Val.getValueType(), VT2); 01415 } 01416 01417 /// Return true if an fneg operation is free to the point where it is never 01418 /// worthwhile to replace it with a bitwise operation. 01419 virtual bool isFNegFree(EVT VT) const { 01420 assert(VT.isFloatingPoint()); 01421 return false; 01422 } 01423 01424 /// Return true if an fabs operation is free to the point where it is never 01425 /// worthwhile to replace it with a bitwise operation. 01426 virtual bool isFAbsFree(EVT VT) const { 01427 assert(VT.isFloatingPoint()); 01428 return false; 01429 } 01430 01431 /// Return true if an FMA operation is faster than a pair of fmul and fadd 01432 /// instructions. fmuladd intrinsics will be expanded to FMAs when this method 01433 /// returns true, otherwise fmuladd is expanded to fmul + fadd. 01434 /// 01435 /// NOTE: This may be called before legalization on types for which FMAs are 01436 /// not legal, but should return true if those types will eventually legalize 01437 /// to types that support FMAs. After legalization, it will only be called on 01438 /// types that support FMAs (via Legal or Custom actions) 01439 virtual bool isFMAFasterThanFMulAndFAdd(EVT) const { 01440 return false; 01441 } 01442 01443 /// Return true if it's profitable to narrow operations of type VT1 to 01444 /// VT2. e.g. on x86, it's profitable to narrow from i32 to i8 but not from 01445 /// i32 to i16. 01446 virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const { 01447 return false; 01448 } 01449 01450 /// \brief Return true if it is beneficial to convert a load of a constant to 01451 /// just the constant itself. 01452 /// On some targets it might be more efficient to use a combination of 01453 /// arithmetic instructions to materialize the constant instead of loading it 01454 /// from a constant pool. 01455 virtual bool shouldConvertConstantLoadToIntImm(const APInt &Imm, 01456 Type *Ty) const { 01457 return false; 01458 } 01459 //===--------------------------------------------------------------------===// 01460 // Runtime Library hooks 01461 // 01462 01463 /// Rename the default libcall routine name for the specified libcall. 01464 void setLibcallName(RTLIB::Libcall Call, const char *Name) { 01465 LibcallRoutineNames[Call] = Name; 01466 } 01467 01468 /// Get the libcall routine name for the specified libcall. 01469 const char *getLibcallName(RTLIB::Libcall Call) const { 01470 return LibcallRoutineNames[Call]; 01471 } 01472 01473 /// Override the default CondCode to be used to test the result of the 01474 /// comparison libcall against zero. 01475 void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) { 01476 CmpLibcallCCs[Call] = CC; 01477 } 01478 01479 /// Get the CondCode that's to be used to test the result of the comparison 01480 /// libcall against zero. 01481 ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const { 01482 return CmpLibcallCCs[Call]; 01483 } 01484 01485 /// Set the CallingConv that should be used for the specified libcall. 01486 void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) { 01487 LibcallCallingConvs[Call] = CC; 01488 } 01489 01490 /// Get the CallingConv that should be used for the specified libcall. 01491 CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const { 01492 return LibcallCallingConvs[Call]; 01493 } 01494 01495 private: 01496 const TargetMachine &TM; 01497 const DataLayout *DL; 01498 const TargetLoweringObjectFile &TLOF; 01499 01500 /// True if this is a little endian target. 01501 bool IsLittleEndian; 01502 01503 /// Tells the code generator not to expand operations into sequences that use 01504 /// the select operations if possible. 01505 bool SelectIsExpensive; 01506 01507 /// Tells the code generator that the target has multiple (allocatable) 01508 /// condition registers that can be used to store the results of comparisons 01509 /// for use by selects and conditional branches. With multiple condition 01510 /// registers, the code generator will not aggressively sink comparisons into 01511 /// the blocks of their users. 01512 bool HasMultipleConditionRegisters; 01513 01514 /// Tells the code generator that the target has BitExtract instructions. 01515 /// The code generator will aggressively sink "shift"s into the blocks of 01516 /// their users if the users will generate "and" instructions which can be 01517 /// combined with "shift" to BitExtract instructions. 01518 bool HasExtractBitsInsn; 01519 01520 /// Tells the code generator not to expand integer divides by constants into a 01521 /// sequence of muls, adds, and shifts. This is a hack until a real cost 01522 /// model is in place. If we ever optimize for size, this will be set to true 01523 /// unconditionally. 01524 bool IntDivIsCheap; 01525 01526 /// Tells the code generator to bypass slow divide or remainder 01527 /// instructions. For example, BypassSlowDivWidths[32,8] tells the code 01528 /// generator to bypass 32-bit integer div/rem with an 8-bit unsigned integer 01529 /// div/rem when the operands are positive and less than 256. 01530 DenseMap <unsigned int, unsigned int> BypassSlowDivWidths; 01531 01532 /// Tells the code generator that it shouldn't generate sra/srl/add/sra for a 01533 /// signed divide by power of two; let the target handle it. 01534 bool Pow2SDivIsCheap; 01535 01536 /// Tells the code generator that it shouldn't generate extra flow control 01537 /// instructions and should attempt to combine flow control instructions via 01538 /// predication. 01539 bool JumpIsExpensive; 01540 01541 /// Whether the target supports or cares about preserving floating point 01542 /// exception behavior. 01543 bool HasFloatingPointExceptions; 01544 01545 /// This target prefers to use _setjmp to implement llvm.setjmp. 01546 /// 01547 /// Defaults to false. 01548 bool UseUnderscoreSetJmp; 01549 01550 /// This target prefers to use _longjmp to implement llvm.longjmp. 01551 /// 01552 /// Defaults to false. 01553 bool UseUnderscoreLongJmp; 01554 01555 /// Number of blocks threshold to use jump tables. 01556 int MinimumJumpTableEntries; 01557 01558 /// Information about the contents of the high-bits in boolean values held in 01559 /// a type wider than i1. See getBooleanContents. 01560 BooleanContent BooleanContents; 01561 01562 /// Information about the contents of the high-bits in boolean values held in 01563 /// a type wider than i1. See getBooleanContents. 01564 BooleanContent BooleanFloatContents; 01565 01566 /// Information about the contents of the high-bits in boolean vector values 01567 /// when the element type is wider than i1. See getBooleanContents. 01568 BooleanContent BooleanVectorContents; 01569 01570 /// The target scheduling preference: shortest possible total cycles or lowest 01571 /// register usage. 01572 Sched::Preference SchedPreferenceInfo; 01573 01574 /// The size, in bytes, of the target's jmp_buf buffers 01575 unsigned JumpBufSize; 01576 01577 /// The alignment, in bytes, of the target's jmp_buf buffers 01578 unsigned JumpBufAlignment; 01579 01580 /// The minimum alignment that any argument on the stack needs to have. 01581 unsigned MinStackArgumentAlignment; 01582 01583 /// The minimum function alignment (used when optimizing for size, and to 01584 /// prevent explicitly provided alignment from leading to incorrect code). 01585 unsigned MinFunctionAlignment; 01586 01587 /// The preferred function alignment (used when alignment unspecified and 01588 /// optimizing for speed). 01589 unsigned PrefFunctionAlignment; 01590 01591 /// The preferred loop alignment. 01592 unsigned PrefLoopAlignment; 01593 01594 /// Whether the DAG builder should automatically insert fences and reduce 01595 /// ordering for atomics. (This will be set for for most architectures with 01596 /// weak memory ordering.) 01597 bool InsertFencesForAtomic; 01598 01599 /// If set to a physical register, this specifies the register that 01600 /// llvm.savestack/llvm.restorestack should save and restore. 01601 unsigned StackPointerRegisterToSaveRestore; 01602 01603 /// If set to a physical register, this specifies the register that receives 01604 /// the exception address on entry to a landing pad. 01605 unsigned ExceptionPointerRegister; 01606 01607 /// If set to a physical register, this specifies the register that receives 01608 /// the exception typeid on entry to a landing pad. 01609 unsigned ExceptionSelectorRegister; 01610 01611 /// This indicates the default register class to use for each ValueType the 01612 /// target supports natively. 01613 const TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE]; 01614 unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE]; 01615 MVT RegisterTypeForVT[MVT::LAST_VALUETYPE]; 01616 01617 /// This indicates the "representative" register class to use for each 01618 /// ValueType the target supports natively. This information is used by the 01619 /// scheduler to track register pressure. By default, the representative 01620 /// register class is the largest legal super-reg register class of the 01621 /// register class of the specified type. e.g. On x86, i8, i16, and i32's 01622 /// representative class would be GR32. 01623 const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE]; 01624 01625 /// This indicates the "cost" of the "representative" register class for each 01626 /// ValueType. The cost is used by the scheduler to approximate register 01627 /// pressure. 01628 uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE]; 01629 01630 /// For any value types we are promoting or expanding, this contains the value 01631 /// type that we are changing to. For Expanded types, this contains one step 01632 /// of the expand (e.g. i64 -> i32), even if there are multiple steps required 01633 /// (e.g. i64 -> i16). For types natively supported by the system, this holds 01634 /// the same type (e.g. i32 -> i32). 01635 MVT TransformToType[MVT::LAST_VALUETYPE]; 01636 01637 /// For each operation and each value type, keep a LegalizeAction that 01638 /// indicates how instruction selection should deal with the operation. Most 01639 /// operations are Legal (aka, supported natively by the target), but 01640 /// operations that are not should be described. Note that operations on 01641 /// non-legal value types are not described here. 01642 uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END]; 01643 01644 /// For each load extension type and each value type, keep a LegalizeAction 01645 /// that indicates how instruction selection should deal with a load of a 01646 /// specific value type and extension type. 01647 uint8_t LoadExtActions[MVT::LAST_VALUETYPE][ISD::LAST_LOADEXT_TYPE]; 01648 01649 /// For each value type pair keep a LegalizeAction that indicates whether a 01650 /// truncating store of a specific value type and truncating type is legal. 01651 uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE]; 01652 01653 /// For each indexed mode and each value type, keep a pair of LegalizeAction 01654 /// that indicates how instruction selection should deal with the load / 01655 /// store. 01656 /// 01657 /// The first dimension is the value_type for the reference. The second 01658 /// dimension represents the various modes for load store. 01659 uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE]; 01660 01661 /// For each condition code (ISD::CondCode) keep a LegalizeAction that 01662 /// indicates how instruction selection should deal with the condition code. 01663 /// 01664 /// Because each CC action takes up 2 bits, we need to have the array size be 01665 /// large enough to fit all of the value types. This can be done by rounding 01666 /// up the MVT::LAST_VALUETYPE value to the next multiple of 16. 01667 uint32_t CondCodeActions[ISD::SETCC_INVALID][(MVT::LAST_VALUETYPE + 15) / 16]; 01668 01669 ValueTypeActionImpl ValueTypeActions; 01670 01671 public: 01672 LegalizeKind 01673 getTypeConversion(LLVMContext &Context, EVT VT) const { 01674 // If this is a simple type, use the ComputeRegisterProp mechanism. 01675 if (VT.isSimple()) { 01676 MVT SVT = VT.getSimpleVT(); 01677 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 01678 MVT NVT = TransformToType[SVT.SimpleTy]; 01679 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 01680 01681 assert( 01682 (LA == TypeLegal || LA == TypeSoftenFloat || 01683 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) 01684 && "Promote may not follow Expand or Promote"); 01685 01686 if (LA == TypeSplitVector) 01687 return LegalizeKind(LA, EVT::getVectorVT(Context, 01688 SVT.getVectorElementType(), 01689 SVT.getVectorNumElements()/2)); 01690 if (LA == TypeScalarizeVector) 01691 return LegalizeKind(LA, SVT.getVectorElementType()); 01692 return LegalizeKind(LA, NVT); 01693 } 01694 01695 // Handle Extended Scalar Types. 01696 if (!VT.isVector()) { 01697 assert(VT.isInteger() && "Float types must be simple"); 01698 unsigned BitSize = VT.getSizeInBits(); 01699 // First promote to a power-of-two size, then expand if necessary. 01700 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 01701 EVT NVT = VT.getRoundIntegerType(Context); 01702 assert(NVT != VT && "Unable to round integer VT"); 01703 LegalizeKind NextStep = getTypeConversion(Context, NVT); 01704 // Avoid multi-step promotion. 01705 if (NextStep.first == TypePromoteInteger) return NextStep; 01706 // Return rounded integer type. 01707 return LegalizeKind(TypePromoteInteger, NVT); 01708 } 01709 01710 return LegalizeKind(TypeExpandInteger, 01711 EVT::getIntegerVT(Context, VT.getSizeInBits()/2)); 01712 } 01713 01714 // Handle vector types. 01715 unsigned NumElts = VT.getVectorNumElements(); 01716 EVT EltVT = VT.getVectorElementType(); 01717 01718 // Vectors with only one element are always scalarized. 01719 if (NumElts == 1) 01720 return LegalizeKind(TypeScalarizeVector, EltVT); 01721 01722 // Try to widen vector elements until the element type is a power of two and 01723 // promote it to a legal type later on, for example: 01724 // <3 x i8> -> <4 x i8> -> <4 x i32> 01725 if (EltVT.isInteger()) { 01726 // Vectors with a number of elements that is not a power of two are always 01727 // widened, for example <3 x i8> -> <4 x i8>. 01728 if (!VT.isPow2VectorType()) { 01729 NumElts = (unsigned)NextPowerOf2(NumElts); 01730 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 01731 return LegalizeKind(TypeWidenVector, NVT); 01732 } 01733 01734 // Examine the element type. 01735 LegalizeKind LK = getTypeConversion(Context, EltVT); 01736 01737 // If type is to be expanded, split the vector. 01738 // <4 x i140> -> <2 x i140> 01739 if (LK.first == TypeExpandInteger) 01740 return LegalizeKind(TypeSplitVector, 01741 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 01742 01743 // Promote the integer element types until a legal vector type is found 01744 // or until the element integer type is too big. If a legal type was not 01745 // found, fallback to the usual mechanism of widening/splitting the 01746 // vector. 01747 EVT OldEltVT = EltVT; 01748 while (1) { 01749 // Increase the bitwidth of the element to the next pow-of-two 01750 // (which is greater than 8 bits). 01751 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits() 01752 ).getRoundIntegerType(Context); 01753 01754 // Stop trying when getting a non-simple element type. 01755 // Note that vector elements may be greater than legal vector element 01756 // types. Example: X86 XMM registers hold 64bit element on 32bit 01757 // systems. 01758 if (!EltVT.isSimple()) break; 01759 01760 // Build a new vector type and check if it is legal. 01761 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 01762 // Found a legal promoted vector type. 01763 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 01764 return LegalizeKind(TypePromoteInteger, 01765 EVT::getVectorVT(Context, EltVT, NumElts)); 01766 } 01767 01768 // Reset the type to the unexpanded type if we did not find a legal vector 01769 // type with a promoted vector element type. 01770 EltVT = OldEltVT; 01771 } 01772 01773 // Try to widen the vector until a legal type is found. 01774 // If there is no wider legal type, split the vector. 01775 while (1) { 01776 // Round up to the next power of 2. 01777 NumElts = (unsigned)NextPowerOf2(NumElts); 01778 01779 // If there is no simple vector type with this many elements then there 01780 // cannot be a larger legal vector type. Note that this assumes that 01781 // there are no skipped intermediate vector types in the simple types. 01782 if (!EltVT.isSimple()) break; 01783 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 01784 if (LargerVector == MVT()) break; 01785 01786 // If this type is legal then widen the vector. 01787 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 01788 return LegalizeKind(TypeWidenVector, LargerVector); 01789 } 01790 01791 // Widen odd vectors to next power of two. 01792 if (!VT.isPow2VectorType()) { 01793 EVT NVT = VT.getPow2VectorType(Context); 01794 return LegalizeKind(TypeWidenVector, NVT); 01795 } 01796 01797 // Vectors with illegal element types are expanded. 01798 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 01799 return LegalizeKind(TypeSplitVector, NVT); 01800 } 01801 01802 private: 01803 std::vector<std::pair<MVT, const TargetRegisterClass*> > AvailableRegClasses; 01804 01805 /// Targets can specify ISD nodes that they would like PerformDAGCombine 01806 /// callbacks for by calling setTargetDAGCombine(), which sets a bit in this 01807 /// array. 01808 unsigned char 01809 TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT]; 01810 01811 /// For operations that must be promoted to a specific type, this holds the 01812 /// destination type. This map should be sparse, so don't hold it as an 01813 /// array. 01814 /// 01815 /// Targets add entries to this map with AddPromotedToType(..), clients access 01816 /// this with getTypeToPromoteTo(..). 01817 std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType> 01818 PromoteToType; 01819 01820 /// Stores the name each libcall. 01821 const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL]; 01822 01823 /// The ISD::CondCode that should be used to test the result of each of the 01824 /// comparison libcall against zero. 01825 ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL]; 01826 01827 /// Stores the CallingConv that should be used for each libcall. 01828 CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL]; 01829 01830 protected: 01831 /// \brief Specify maximum number of store instructions per memset call. 01832 /// 01833 /// When lowering \@llvm.memset this field specifies the maximum number of 01834 /// store operations that may be substituted for the call to memset. Targets 01835 /// must set this value based on the cost threshold for that target. Targets 01836 /// should assume that the memset will be done using as many of the largest 01837 /// store operations first, followed by smaller ones, if necessary, per 01838 /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine 01839 /// with 16-bit alignment would result in four 2-byte stores and one 1-byte 01840 /// store. This only applies to setting a constant array of a constant size. 01841 unsigned MaxStoresPerMemset; 01842 01843 /// Maximum number of stores operations that may be substituted for the call 01844 /// to memset, used for functions with OptSize attribute. 01845 unsigned MaxStoresPerMemsetOptSize; 01846 01847 /// \brief Specify maximum bytes of store instructions per memcpy call. 01848 /// 01849 /// When lowering \@llvm.memcpy this field specifies the maximum number of 01850 /// store operations that may be substituted for a call to memcpy. Targets 01851 /// must set this value based on the cost threshold for that target. Targets 01852 /// should assume that the memcpy will be done using as many of the largest 01853 /// store operations first, followed by smaller ones, if necessary, per 01854 /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine 01855 /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store 01856 /// and one 1-byte store. This only applies to copying a constant array of 01857 /// constant size. 01858 unsigned MaxStoresPerMemcpy; 01859 01860 /// Maximum number of store operations that may be substituted for a call to 01861 /// memcpy, used for functions with OptSize attribute. 01862 unsigned MaxStoresPerMemcpyOptSize; 01863 01864 /// \brief Specify maximum bytes of store instructions per memmove call. 01865 /// 01866 /// When lowering \@llvm.memmove this field specifies the maximum number of 01867 /// store instructions that may be substituted for a call to memmove. Targets 01868 /// must set this value based on the cost threshold for that target. Targets 01869 /// should assume that the memmove will be done using as many of the largest 01870 /// store operations first, followed by smaller ones, if necessary, per 01871 /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine 01872 /// with 8-bit alignment would result in nine 1-byte stores. This only 01873 /// applies to copying a constant array of constant size. 01874 unsigned MaxStoresPerMemmove; 01875 01876 /// Maximum number of store instructions that may be substituted for a call to 01877 /// memmove, used for functions with OpSize attribute. 01878 unsigned MaxStoresPerMemmoveOptSize; 01879 01880 /// Tells the code generator that select is more expensive than a branch if 01881 /// the branch is usually predicted right. 01882 bool PredictableSelectIsExpensive; 01883 01884 /// MaskAndBranchFoldingIsLegal - Indicates if the target supports folding 01885 /// a mask of a single bit, a compare, and a branch into a single instruction. 01886 bool MaskAndBranchFoldingIsLegal; 01887 01888 protected: 01889 /// Return true if the value types that can be represented by the specified 01890 /// register class are all legal. 01891 bool isLegalRC(const TargetRegisterClass *RC) const; 01892 01893 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 01894 /// sequence of memory operands that is recognized by PrologEpilogInserter. 01895 MachineBasicBlock *emitPatchPoint(MachineInstr *MI, MachineBasicBlock *MBB) const; 01896 }; 01897 01898 /// This class defines information used to lower LLVM code to legal SelectionDAG 01899 /// operators that the target instruction selector can accept natively. 01900 /// 01901 /// This class also defines callbacks that targets must implement to lower 01902 /// target-specific constructs to SelectionDAG operators. 01903 class TargetLowering : public TargetLoweringBase { 01904 TargetLowering(const TargetLowering&) LLVM_DELETED_FUNCTION; 01905 void operator=(const TargetLowering&) LLVM_DELETED_FUNCTION; 01906 01907 public: 01908 /// NOTE: The constructor takes ownership of TLOF. 01909 explicit TargetLowering(const TargetMachine &TM, 01910 const TargetLoweringObjectFile *TLOF); 01911 01912 /// Returns true by value, base pointer and offset pointer and addressing mode 01913 /// by reference if the node's address can be legally represented as 01914 /// pre-indexed load / store address. 01915 virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/, 01916 SDValue &/*Offset*/, 01917 ISD::MemIndexedMode &/*AM*/, 01918 SelectionDAG &/*DAG*/) const { 01919 return false; 01920 } 01921 01922 /// Returns true by value, base pointer and offset pointer and addressing mode 01923 /// by reference if this node can be combined with a load / store to form a 01924 /// post-indexed load / store. 01925 virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/, 01926 SDValue &/*Base*/, 01927 SDValue &/*Offset*/, 01928 ISD::MemIndexedMode &/*AM*/, 01929 SelectionDAG &/*DAG*/) const { 01930 return false; 01931 } 01932 01933 /// Return the entry encoding for a jump table in the current function. The 01934 /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum. 01935 virtual unsigned getJumpTableEncoding() const; 01936 01937 virtual const MCExpr * 01938 LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/, 01939 const MachineBasicBlock * /*MBB*/, unsigned /*uid*/, 01940 MCContext &/*Ctx*/) const { 01941 llvm_unreachable("Need to implement this hook if target has custom JTIs"); 01942 } 01943 01944 /// Returns relocation base for the given PIC jumptable. 01945 virtual SDValue getPICJumpTableRelocBase(SDValue Table, 01946 SelectionDAG &DAG) const; 01947 01948 /// This returns the relocation base for the given PIC jumptable, the same as 01949 /// getPICJumpTableRelocBase, but as an MCExpr. 01950 virtual const MCExpr * 01951 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 01952 unsigned JTI, MCContext &Ctx) const; 01953 01954 /// Return true if folding a constant offset with the given GlobalAddress is 01955 /// legal. It is frequently not legal in PIC relocation models. 01956 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const; 01957 01958 bool isInTailCallPosition(SelectionDAG &DAG, SDNode *Node, 01959 SDValue &Chain) const; 01960 01961 void softenSetCCOperands(SelectionDAG &DAG, EVT VT, 01962 SDValue &NewLHS, SDValue &NewRHS, 01963 ISD::CondCode &CCCode, SDLoc DL) const; 01964 01965 /// Returns a pair of (return value, chain). 01966 std::pair<SDValue, SDValue> makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC, 01967 EVT RetVT, const SDValue *Ops, 01968 unsigned NumOps, bool isSigned, 01969 SDLoc dl, bool doesNotReturn = false, 01970 bool isReturnValueUsed = true) const; 01971 01972 //===--------------------------------------------------------------------===// 01973 // TargetLowering Optimization Methods 01974 // 01975 01976 /// A convenience struct that encapsulates a DAG, and two SDValues for 01977 /// returning information from TargetLowering to its clients that want to 01978 /// combine. 01979 struct TargetLoweringOpt { 01980 SelectionDAG &DAG; 01981 bool LegalTys; 01982 bool LegalOps; 01983 SDValue Old; 01984 SDValue New; 01985 01986 explicit TargetLoweringOpt(SelectionDAG &InDAG, 01987 bool LT, bool LO) : 01988 DAG(InDAG), LegalTys(LT), LegalOps(LO) {} 01989 01990 bool LegalTypes() const { return LegalTys; } 01991 bool LegalOperations() const { return LegalOps; } 01992 01993 bool CombineTo(SDValue O, SDValue N) { 01994 Old = O; 01995 New = N; 01996 return true; 01997 } 01998 01999 /// Check to see if the specified operand of the specified instruction is a 02000 /// constant integer. If so, check to see if there are any bits set in the 02001 /// constant that are not demanded. If so, shrink the constant and return 02002 /// true. 02003 bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded); 02004 02005 /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free. This 02006 /// uses isZExtFree and ZERO_EXTEND for the widening cast, but it could be 02007 /// generalized for targets with other types of implicit widening casts. 02008 bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded, 02009 SDLoc dl); 02010 }; 02011 02012 /// Look at Op. At this point, we know that only the DemandedMask bits of the 02013 /// result of Op are ever used downstream. If we can use this information to 02014 /// simplify Op, create a new simplified DAG node and return true, returning 02015 /// the original and new nodes in Old and New. Otherwise, analyze the 02016 /// expression and return a mask of KnownOne and KnownZero bits for the 02017 /// expression (used to simplify the caller). The KnownZero/One bits may only 02018 /// be accurate for those bits in the DemandedMask. 02019 bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask, 02020 APInt &KnownZero, APInt &KnownOne, 02021 TargetLoweringOpt &TLO, unsigned Depth = 0) const; 02022 02023 /// Determine which of the bits specified in Mask are known to be either zero 02024 /// or one and return them in the KnownZero/KnownOne bitsets. 02025 virtual void computeKnownBitsForTargetNode(const SDValue Op, 02026 APInt &KnownZero, 02027 APInt &KnownOne, 02028 const SelectionDAG &DAG, 02029 unsigned Depth = 0) const; 02030 02031 /// This method can be implemented by targets that want to expose additional 02032 /// information about sign bits to the DAG Combiner. 02033 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op, 02034 const SelectionDAG &DAG, 02035 unsigned Depth = 0) const; 02036 02037 struct DAGCombinerInfo { 02038 void *DC; // The DAG Combiner object. 02039 CombineLevel Level; 02040 bool CalledByLegalizer; 02041 public: 02042 SelectionDAG &DAG; 02043 02044 DAGCombinerInfo(SelectionDAG &dag, CombineLevel level, bool cl, void *dc) 02045 : DC(dc), Level(level), CalledByLegalizer(cl), DAG(dag) {} 02046 02047 bool isBeforeLegalize() const { return Level == BeforeLegalizeTypes; } 02048 bool isBeforeLegalizeOps() const { return Level < AfterLegalizeVectorOps; } 02049 bool isAfterLegalizeVectorOps() const { 02050 return Level == AfterLegalizeDAG; 02051 } 02052 CombineLevel getDAGCombineLevel() { return Level; } 02053 bool isCalledByLegalizer() const { return CalledByLegalizer; } 02054 02055 void AddToWorklist(SDNode *N); 02056 void RemoveFromWorklist(SDNode *N); 02057 SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To, 02058 bool AddTo = true); 02059 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true); 02060 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true); 02061 02062 void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO); 02063 }; 02064 02065 /// Return if the N is a constant or constant vector equal to the true value 02066 /// from getBooleanContents(). 02067 bool isConstTrueVal(const SDNode *N) const; 02068 02069 /// Return if the N is a constant or constant vector equal to the false value 02070 /// from getBooleanContents(). 02071 bool isConstFalseVal(const SDNode *N) const; 02072 02073 /// Try to simplify a setcc built with the specified operands and cc. If it is 02074 /// unable to simplify it, return a null SDValue. 02075 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 02076 ISD::CondCode Cond, bool foldBooleans, 02077 DAGCombinerInfo &DCI, SDLoc dl) const; 02078 02079 /// Returns true (and the GlobalValue and the offset) if the node is a 02080 /// GlobalAddress + offset. 02081 virtual bool 02082 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const; 02083 02084 /// This method will be invoked for all target nodes and for any 02085 /// target-independent nodes that the target has registered with invoke it 02086 /// for. 02087 /// 02088 /// The semantics are as follows: 02089 /// Return Value: 02090 /// SDValue.Val == 0 - No change was made 02091 /// SDValue.Val == N - N was replaced, is dead, and is already handled. 02092 /// otherwise - N should be replaced by the returned Operand. 02093 /// 02094 /// In addition, methods provided by DAGCombinerInfo may be used to perform 02095 /// more complex transformations. 02096 /// 02097 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 02098 02099 /// Return true if it is profitable to move a following shift through this 02100 // node, adjusting any immediate operands as necessary to preserve semantics. 02101 // This transformation may not be desirable if it disrupts a particularly 02102 // auspicious target-specific tree (e.g. bitfield extraction in AArch64). 02103 // By default, it returns true. 02104 virtual bool isDesirableToCommuteWithShift(const SDNode *N /*Op*/) const { 02105 return true; 02106 } 02107 02108 /// Return true if the target has native support for the specified value type 02109 /// and it is 'desirable' to use the type for the given node type. e.g. On x86 02110 /// i16 is legal, but undesirable since i16 instruction encodings are longer 02111 /// and some i16 instructions are slow. 02112 virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const { 02113 // By default, assume all legal types are desirable. 02114 return isTypeLegal(VT); 02115 } 02116 02117 /// Return true if it is profitable for dag combiner to transform a floating 02118 /// point op of specified opcode to a equivalent op of an integer 02119 /// type. e.g. f32 load -> i32 load can be profitable on ARM. 02120 virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/, 02121 EVT /*VT*/) const { 02122 return false; 02123 } 02124 02125 /// This method query the target whether it is beneficial for dag combiner to 02126 /// promote the specified node. If true, it should return the desired 02127 /// promotion type by reference. 02128 virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const { 02129 return false; 02130 } 02131 02132 //===--------------------------------------------------------------------===// 02133 // Lowering methods - These methods must be implemented by targets so that 02134 // the SelectionDAGBuilder code knows how to lower these. 02135 // 02136 02137 /// This hook must be implemented to lower the incoming (formal) arguments, 02138 /// described by the Ins array, into the specified DAG. The implementation 02139 /// should fill in the InVals array with legal-type argument values, and 02140 /// return the resulting token chain value. 02141 /// 02142 virtual SDValue 02143 LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 02144 bool /*isVarArg*/, 02145 const SmallVectorImpl<ISD::InputArg> &/*Ins*/, 02146 SDLoc /*dl*/, SelectionDAG &/*DAG*/, 02147 SmallVectorImpl<SDValue> &/*InVals*/) const { 02148 llvm_unreachable("Not Implemented"); 02149 } 02150 02151 struct ArgListEntry { 02152 SDValue Node; 02153 Type* Ty; 02154 bool isSExt : 1; 02155 bool isZExt : 1; 02156 bool isInReg : 1; 02157 bool isSRet : 1; 02158 bool isNest : 1; 02159 bool isByVal : 1; 02160 bool isInAlloca : 1; 02161 bool isReturned : 1; 02162 uint16_t Alignment; 02163 02164 ArgListEntry() : isSExt(false), isZExt(false), isInReg(false), 02165 isSRet(false), isNest(false), isByVal(false), isInAlloca(false), 02166 isReturned(false), Alignment(0) { } 02167 02168 void setAttributes(ImmutableCallSite *CS, unsigned AttrIdx); 02169 }; 02170 typedef std::vector<ArgListEntry> ArgListTy; 02171 02172 /// This structure contains all information that is necessary for lowering 02173 /// calls. It is passed to TLI::LowerCallTo when the SelectionDAG builder 02174 /// needs to lower a call, and targets will see this struct in their LowerCall 02175 /// implementation. 02176 struct CallLoweringInfo { 02177 SDValue Chain; 02178 Type *RetTy; 02179 bool RetSExt : 1; 02180 bool RetZExt : 1; 02181 bool IsVarArg : 1; 02182 bool IsInReg : 1; 02183 bool DoesNotReturn : 1; 02184 bool IsReturnValueUsed : 1; 02185 02186 // IsTailCall should be modified by implementations of 02187 // TargetLowering::LowerCall that perform tail call conversions. 02188 bool IsTailCall; 02189 02190 unsigned NumFixedArgs; 02191 CallingConv::ID CallConv; 02192 SDValue Callee; 02193 ArgListTy Args; 02194 SelectionDAG &DAG; 02195 SDLoc DL; 02196 ImmutableCallSite *CS; 02197 SmallVector<ISD::OutputArg, 32> Outs; 02198 SmallVector<SDValue, 32> OutVals; 02199 SmallVector<ISD::InputArg, 32> Ins; 02200 02201 CallLoweringInfo(SelectionDAG &DAG) 02202 : RetTy(nullptr), RetSExt(false), RetZExt(false), IsVarArg(false), 02203 IsInReg(false), DoesNotReturn(false), IsReturnValueUsed(true), 02204 IsTailCall(false), NumFixedArgs(-1), CallConv(CallingConv::C), 02205 DAG(DAG), CS(nullptr) {} 02206 02207 CallLoweringInfo &setDebugLoc(SDLoc dl) { 02208 DL = dl; 02209 return *this; 02210 } 02211 02212 CallLoweringInfo &setChain(SDValue InChain) { 02213 Chain = InChain; 02214 return *this; 02215 } 02216 02217 CallLoweringInfo &setCallee(CallingConv::ID CC, Type *ResultType, 02218 SDValue Target, ArgListTy &&ArgsList, 02219 unsigned FixedArgs = -1) { 02220 RetTy = ResultType; 02221 Callee = Target; 02222 CallConv = CC; 02223 NumFixedArgs = 02224 (FixedArgs == static_cast<unsigned>(-1) ? Args.size() : FixedArgs); 02225 Args = std::move(ArgsList); 02226 return *this; 02227 } 02228 02229 CallLoweringInfo &setCallee(Type *ResultType, FunctionType *FTy, 02230 SDValue Target, ArgListTy &&ArgsList, 02231 ImmutableCallSite &Call) { 02232 RetTy = ResultType; 02233 02234 IsInReg = Call.paramHasAttr(0, Attribute::InReg); 02235 DoesNotReturn = Call.doesNotReturn(); 02236 IsVarArg = FTy->isVarArg(); 02237 IsReturnValueUsed = !Call.getInstruction()->use_empty(); 02238 RetSExt = Call.paramHasAttr(0, Attribute::SExt); 02239 RetZExt = Call.paramHasAttr(0, Attribute::ZExt); 02240 02241 Callee = Target; 02242 02243 CallConv = Call.getCallingConv(); 02244 NumFixedArgs = FTy->getNumParams(); 02245 Args = std::move(ArgsList); 02246 02247 CS = &Call; 02248 02249 return *this; 02250 } 02251 02252 CallLoweringInfo &setInRegister(bool Value = true) { 02253 IsInReg = Value; 02254 return *this; 02255 } 02256 02257 CallLoweringInfo &setNoReturn(bool Value = true) { 02258 DoesNotReturn = Value; 02259 return *this; 02260 } 02261 02262 CallLoweringInfo &setVarArg(bool Value = true) { 02263 IsVarArg = Value; 02264 return *this; 02265 } 02266 02267 CallLoweringInfo &setTailCall(bool Value = true) { 02268 IsTailCall = Value; 02269 return *this; 02270 } 02271 02272 CallLoweringInfo &setDiscardResult(bool Value = true) { 02273 IsReturnValueUsed = !Value; 02274 return *this; 02275 } 02276 02277 CallLoweringInfo &setSExtResult(bool Value = true) { 02278 RetSExt = Value; 02279 return *this; 02280 } 02281 02282 CallLoweringInfo &setZExtResult(bool Value = true) { 02283 RetZExt = Value; 02284 return *this; 02285 } 02286 02287 ArgListTy &getArgs() { 02288 return Args; 02289 } 02290 }; 02291 02292 /// This function lowers an abstract call to a function into an actual call. 02293 /// This returns a pair of operands. The first element is the return value 02294 /// for the function (if RetTy is not VoidTy). The second element is the 02295 /// outgoing token chain. It calls LowerCall to do the actual lowering. 02296 std::pair<SDValue, SDValue> LowerCallTo(CallLoweringInfo &CLI) const; 02297 02298 /// This hook must be implemented to lower calls into the the specified 02299 /// DAG. The outgoing arguments to the call are described by the Outs array, 02300 /// and the values to be returned by the call are described by the Ins 02301 /// array. The implementation should fill in the InVals array with legal-type 02302 /// return values from the call, and return the resulting token chain value. 02303 virtual SDValue 02304 LowerCall(CallLoweringInfo &/*CLI*/, 02305 SmallVectorImpl<SDValue> &/*InVals*/) const { 02306 llvm_unreachable("Not Implemented"); 02307 } 02308 02309 /// Target-specific cleanup for formal ByVal parameters. 02310 virtual void HandleByVal(CCState *, unsigned &, unsigned) const {} 02311 02312 /// This hook should be implemented to check whether the return values 02313 /// described by the Outs array can fit into the return registers. If false 02314 /// is returned, an sret-demotion is performed. 02315 virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/, 02316 MachineFunction &/*MF*/, bool /*isVarArg*/, 02317 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 02318 LLVMContext &/*Context*/) const 02319 { 02320 // Return true by default to get preexisting behavior. 02321 return true; 02322 } 02323 02324 /// This hook must be implemented to lower outgoing return values, described 02325 /// by the Outs array, into the specified DAG. The implementation should 02326 /// return the resulting token chain value. 02327 virtual SDValue 02328 LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 02329 bool /*isVarArg*/, 02330 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 02331 const SmallVectorImpl<SDValue> &/*OutVals*/, 02332 SDLoc /*dl*/, SelectionDAG &/*DAG*/) const { 02333 llvm_unreachable("Not Implemented"); 02334 } 02335 02336 /// Return true if result of the specified node is used by a return node 02337 /// only. It also compute and return the input chain for the tail call. 02338 /// 02339 /// This is used to determine whether it is possible to codegen a libcall as 02340 /// tail call at legalization time. 02341 virtual bool isUsedByReturnOnly(SDNode *, SDValue &/*Chain*/) const { 02342 return false; 02343 } 02344 02345 /// Return true if the target may be able emit the call instruction as a tail 02346 /// call. This is used by optimization passes to determine if it's profitable 02347 /// to duplicate return instructions to enable tailcall optimization. 02348 virtual bool mayBeEmittedAsTailCall(CallInst *) const { 02349 return false; 02350 } 02351 02352 /// Return the builtin name for the __builtin___clear_cache intrinsic 02353 /// Default is to invoke the clear cache library call 02354 virtual const char * getClearCacheBuiltinName() const { 02355 return "__clear_cache"; 02356 } 02357 02358 /// Return the register ID of the name passed in. Used by named register 02359 /// global variables extension. There is no target-independent behaviour 02360 /// so the default action is to bail. 02361 virtual unsigned getRegisterByName(const char* RegName, EVT VT) const { 02362 report_fatal_error("Named registers not implemented for this target"); 02363 } 02364 02365 /// Return the type that should be used to zero or sign extend a 02366 /// zeroext/signext integer argument or return value. FIXME: Most C calling 02367 /// convention requires the return type to be promoted, but this is not true 02368 /// all the time, e.g. i1 on x86-64. It is also not necessary for non-C 02369 /// calling conventions. The frontend should handle this and include all of 02370 /// the necessary information. 02371 virtual EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT, 02372 ISD::NodeType /*ExtendKind*/) const { 02373 EVT MinVT = getRegisterType(Context, MVT::i32); 02374 return VT.bitsLT(MinVT) ? MinVT : VT; 02375 } 02376 02377 /// For some targets, an LLVM struct type must be broken down into multiple 02378 /// simple types, but the calling convention specifies that the entire struct 02379 /// must be passed in a block of consecutive registers. 02380 virtual bool 02381 functionArgumentNeedsConsecutiveRegisters(Type *Ty, CallingConv::ID CallConv, 02382 bool isVarArg) const { 02383 return false; 02384 } 02385 02386 /// Returns a 0 terminated array of registers that can be safely used as 02387 /// scratch registers. 02388 virtual const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const { 02389 return nullptr; 02390 } 02391 02392 /// This callback is used to prepare for a volatile or atomic load. 02393 /// It takes a chain node as input and returns the chain for the load itself. 02394 /// 02395 /// Having a callback like this is necessary for targets like SystemZ, 02396 /// which allows a CPU to reuse the result of a previous load indefinitely, 02397 /// even if a cache-coherent store is performed by another CPU. The default 02398 /// implementation does nothing. 02399 virtual SDValue prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, 02400 SelectionDAG &DAG) const { 02401 return Chain; 02402 } 02403 02404 /// This callback is invoked by the type legalizer to legalize nodes with an 02405 /// illegal operand type but legal result types. It replaces the 02406 /// LowerOperation callback in the type Legalizer. The reason we can not do 02407 /// away with LowerOperation entirely is that LegalizeDAG isn't yet ready to 02408 /// use this callback. 02409 /// 02410 /// TODO: Consider merging with ReplaceNodeResults. 02411 /// 02412 /// The target places new result values for the node in Results (their number 02413 /// and types must exactly match those of the original return values of 02414 /// the node), or leaves Results empty, which indicates that the node is not 02415 /// to be custom lowered after all. 02416 /// The default implementation calls LowerOperation. 02417 virtual void LowerOperationWrapper(SDNode *N, 02418 SmallVectorImpl<SDValue> &Results, 02419 SelectionDAG &DAG) const; 02420 02421 /// This callback is invoked for operations that are unsupported by the 02422 /// target, which are registered to use 'custom' lowering, and whose defined 02423 /// values are all legal. If the target has no operations that require custom 02424 /// lowering, it need not implement this. The default implementation of this 02425 /// aborts. 02426 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; 02427 02428 /// This callback is invoked when a node result type is illegal for the 02429 /// target, and the operation was registered to use 'custom' lowering for that 02430 /// result type. The target places new result values for the node in Results 02431 /// (their number and types must exactly match those of the original return 02432 /// values of the node), or leaves Results empty, which indicates that the 02433 /// node is not to be custom lowered after all. 02434 /// 02435 /// If the target has no operations that require custom lowering, it need not 02436 /// implement this. The default implementation aborts. 02437 virtual void ReplaceNodeResults(SDNode * /*N*/, 02438 SmallVectorImpl<SDValue> &/*Results*/, 02439 SelectionDAG &/*DAG*/) const { 02440 llvm_unreachable("ReplaceNodeResults not implemented for this target!"); 02441 } 02442 02443 /// This method returns the name of a target specific DAG node. 02444 virtual const char *getTargetNodeName(unsigned Opcode) const; 02445 02446 /// This method returns a target specific FastISel object, or null if the 02447 /// target does not support "fast" ISel. 02448 virtual FastISel *createFastISel(FunctionLoweringInfo &, 02449 const TargetLibraryInfo *) const { 02450 return nullptr; 02451 } 02452 02453 02454 bool verifyReturnAddressArgumentIsConstant(SDValue Op, 02455 SelectionDAG &DAG) const; 02456 02457 //===--------------------------------------------------------------------===// 02458 // Inline Asm Support hooks 02459 // 02460 02461 /// This hook allows the target to expand an inline asm call to be explicit 02462 /// llvm code if it wants to. This is useful for turning simple inline asms 02463 /// into LLVM intrinsics, which gives the compiler more information about the 02464 /// behavior of the code. 02465 virtual bool ExpandInlineAsm(CallInst *) const { 02466 return false; 02467 } 02468 02469 enum ConstraintType { 02470 C_Register, // Constraint represents specific register(s). 02471 C_RegisterClass, // Constraint represents any of register(s) in class. 02472 C_Memory, // Memory constraint. 02473 C_Other, // Something else. 02474 C_Unknown // Unsupported constraint. 02475 }; 02476 02477 enum ConstraintWeight { 02478 // Generic weights. 02479 CW_Invalid = -1, // No match. 02480 CW_Okay = 0, // Acceptable. 02481 CW_Good = 1, // Good weight. 02482 CW_Better = 2, // Better weight. 02483 CW_Best = 3, // Best weight. 02484 02485 // Well-known weights. 02486 CW_SpecificReg = CW_Okay, // Specific register operands. 02487 CW_Register = CW_Good, // Register operands. 02488 CW_Memory = CW_Better, // Memory operands. 02489 CW_Constant = CW_Best, // Constant operand. 02490 CW_Default = CW_Okay // Default or don't know type. 02491 }; 02492 02493 /// This contains information for each constraint that we are lowering. 02494 struct AsmOperandInfo : public InlineAsm::ConstraintInfo { 02495 /// This contains the actual string for the code, like "m". TargetLowering 02496 /// picks the 'best' code from ConstraintInfo::Codes that most closely 02497 /// matches the operand. 02498 std::string ConstraintCode; 02499 02500 /// Information about the constraint code, e.g. Register, RegisterClass, 02501 /// Memory, Other, Unknown. 02502 TargetLowering::ConstraintType ConstraintType; 02503 02504 /// If this is the result output operand or a clobber, this is null, 02505 /// otherwise it is the incoming operand to the CallInst. This gets 02506 /// modified as the asm is processed. 02507 Value *CallOperandVal; 02508 02509 /// The ValueType for the operand value. 02510 MVT ConstraintVT; 02511 02512 /// Return true of this is an input operand that is a matching constraint 02513 /// like "4". 02514 bool isMatchingInputConstraint() const; 02515 02516 /// If this is an input matching constraint, this method returns the output 02517 /// operand it matches. 02518 unsigned getMatchedOperand() const; 02519 02520 /// Copy constructor for copying from a ConstraintInfo. 02521 AsmOperandInfo(const InlineAsm::ConstraintInfo &info) 02522 : InlineAsm::ConstraintInfo(info), 02523 ConstraintType(TargetLowering::C_Unknown), 02524 CallOperandVal(nullptr), ConstraintVT(MVT::Other) { 02525 } 02526 }; 02527 02528 typedef std::vector<AsmOperandInfo> AsmOperandInfoVector; 02529 02530 /// Split up the constraint string from the inline assembly value into the 02531 /// specific constraints and their prefixes, and also tie in the associated 02532 /// operand values. If this returns an empty vector, and if the constraint 02533 /// string itself isn't empty, there was an error parsing. 02534 virtual AsmOperandInfoVector ParseConstraints(ImmutableCallSite CS) const; 02535 02536 /// Examine constraint type and operand type and determine a weight value. 02537 /// The operand object must already have been set up with the operand type. 02538 virtual ConstraintWeight getMultipleConstraintMatchWeight( 02539 AsmOperandInfo &info, int maIndex) const; 02540 02541 /// Examine constraint string and operand type and determine a weight value. 02542 /// The operand object must already have been set up with the operand type. 02543 virtual ConstraintWeight getSingleConstraintMatchWeight( 02544 AsmOperandInfo &info, const char *constraint) const; 02545 02546 /// Determines the constraint code and constraint type to use for the specific 02547 /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType. 02548 /// If the actual operand being passed in is available, it can be passed in as 02549 /// Op, otherwise an empty SDValue can be passed. 02550 virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo, 02551 SDValue Op, 02552 SelectionDAG *DAG = nullptr) const; 02553 02554 /// Given a constraint, return the type of constraint it is for this target. 02555 virtual ConstraintType getConstraintType(const std::string &Constraint) const; 02556 02557 /// Given a physical register constraint (e.g. {edx}), return the register 02558 /// number and the register class for the register. 02559 /// 02560 /// Given a register class constraint, like 'r', if this corresponds directly 02561 /// to an LLVM register class, return a register of 0 and the register class 02562 /// pointer. 02563 /// 02564 /// This should only be used for C_Register constraints. On error, this 02565 /// returns a register number of 0 and a null register class pointer.. 02566 virtual std::pair<unsigned, const TargetRegisterClass*> 02567 getRegForInlineAsmConstraint(const std::string &Constraint, 02568 MVT VT) const; 02569 02570 /// Try to replace an X constraint, which matches anything, with another that 02571 /// has more specific requirements based on the type of the corresponding 02572 /// operand. This returns null if there is no replacement to make. 02573 virtual const char *LowerXConstraint(EVT ConstraintVT) const; 02574 02575 /// Lower the specified operand into the Ops vector. If it is invalid, don't 02576 /// add anything to Ops. 02577 virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 02578 std::vector<SDValue> &Ops, 02579 SelectionDAG &DAG) const; 02580 02581 //===--------------------------------------------------------------------===// 02582 // Div utility functions 02583 // 02584 SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl, 02585 SelectionDAG &DAG) const; 02586 SDValue BuildSDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 02587 bool IsAfterLegalization, 02588 std::vector<SDNode *> *Created) const; 02589 SDValue BuildUDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 02590 bool IsAfterLegalization, 02591 std::vector<SDNode *> *Created) const; 02592 virtual SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, 02593 SelectionDAG &DAG, 02594 std::vector<SDNode *> *Created) const { 02595 return SDValue(); 02596 } 02597 02598 //===--------------------------------------------------------------------===// 02599 // Legalization utility functions 02600 // 02601 02602 /// Expand a MUL into two nodes. One that computes the high bits of 02603 /// the result and one that computes the low bits. 02604 /// \param HiLoVT The value type to use for the Lo and Hi nodes. 02605 /// \param LL Low bits of the LHS of the MUL. You can use this parameter 02606 /// if you want to control how low bits are extracted from the LHS. 02607 /// \param LH High bits of the LHS of the MUL. See LL for meaning. 02608 /// \param RL Low bits of the RHS of the MUL. See LL for meaning 02609 /// \param RH High bits of the RHS of the MUL. See LL for meaning. 02610 /// \returns true if the node has been expanded. false if it has not 02611 bool expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT, 02612 SelectionDAG &DAG, SDValue LL = SDValue(), 02613 SDValue LH = SDValue(), SDValue RL = SDValue(), 02614 SDValue RH = SDValue()) const; 02615 02616 /// Expand float(f32) to SINT(i64) conversion 02617 /// \param N Node to expand 02618 /// \param Result output after conversion 02619 /// \returns True, if the expansion was successful, false otherwise 02620 bool expandFP_TO_SINT(SDNode *N, SDValue &Result, SelectionDAG &DAG) const; 02621 02622 //===--------------------------------------------------------------------===// 02623 // Instruction Emitting Hooks 02624 // 02625 02626 /// This method should be implemented by targets that mark instructions with 02627 /// the 'usesCustomInserter' flag. These instructions are special in various 02628 /// ways, which require special support to insert. The specified MachineInstr 02629 /// is created but not inserted into any basic blocks, and this method is 02630 /// called to expand it into a sequence of instructions, potentially also 02631 /// creating new basic blocks and control flow. 02632 virtual MachineBasicBlock * 02633 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const; 02634 02635 /// This method should be implemented by targets that mark instructions with 02636 /// the 'hasPostISelHook' flag. These instructions must be adjusted after 02637 /// instruction selection by target hooks. e.g. To fill in optional defs for 02638 /// ARM 's' setting instructions. 02639 virtual void 02640 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const; 02641 02642 /// If this function returns true, SelectionDAGBuilder emits a 02643 /// LOAD_STACK_GUARD node when it is lowering Intrinsic::stackprotector. 02644 virtual bool useLoadStackGuardNode() const { 02645 return false; 02646 } 02647 }; 02648 02649 /// Given an LLVM IR type and return type attributes, compute the return value 02650 /// EVTs and flags, and optionally also the offsets, if the return value is 02651 /// being lowered to memory. 02652 void GetReturnInfo(Type* ReturnType, AttributeSet attr, 02653 SmallVectorImpl<ISD::OutputArg> &Outs, 02654 const TargetLowering &TLI); 02655 02656 } // end llvm namespace 02657 02658 #endif