1 //===-- llvm/Target/TargetLowering.h - Target Lowering Info -----*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// This file describes how to lower LLVM code to machine code. This has two
14 /// 1. Which ValueTypes are natively supported by the target.
15 /// 2. Which operations are supported for supported ValueTypes.
16 /// 3. Cost thresholds for alternative implementations of certain operations.
18 /// In addition it has a few other components, like information about FP
21 //===----------------------------------------------------------------------===//
23 #ifndef LLVM_TARGET_TARGETLOWERING_H
24 #define LLVM_TARGET_TARGETLOWERING_H
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/CodeGen/DAGCombine.h"
28 #include "llvm/CodeGen/RuntimeLibcalls.h"
29 #include "llvm/CodeGen/SelectionDAGNodes.h"
30 #include "llvm/IR/Attributes.h"
31 #include "llvm/IR/CallSite.h"
32 #include "llvm/IR/CallingConv.h"
33 #include "llvm/IR/InlineAsm.h"
34 #include "llvm/IR/IRBuilder.h"
35 #include "llvm/IR/GlobalVariable.h"
36 #include "llvm/MC/MCRegisterInfo.h"
37 #include "llvm/Target/TargetCallingConv.h"
38 #include "llvm/Target/TargetMachine.h"
47 class FunctionLoweringInfo;
48 class ImmutableCallSite;
50 class MachineBasicBlock;
51 class MachineFunction;
53 class MachineJumpTableInfo;
58 template<typename T> class SmallVectorImpl;
60 class TargetRegisterClass;
61 class TargetLibraryInfo;
62 class TargetLoweringObjectFile;
67 None, // No preference
68 Source, // Follow source order.
69 RegPressure, // Scheduling for lowest register pressure.
70 Hybrid, // Scheduling for both latency and register pressure.
71 ILP, // Scheduling for ILP in low register pressure mode.
72 VLIW // Scheduling for VLIW targets.
76 /// This base class for TargetLowering contains the SelectionDAG-independent
77 /// parts that can be used from the rest of CodeGen.
78 class TargetLoweringBase {
79 TargetLoweringBase(const TargetLoweringBase&) LLVM_DELETED_FUNCTION;
80 void operator=(const TargetLoweringBase&) LLVM_DELETED_FUNCTION;
83 /// This enum indicates whether operations are valid for a target, and if not,
84 /// what action should be used to make them valid.
86 Legal, // The target natively supports this operation.
87 Promote, // This operation should be executed in a larger type.
88 Expand, // Try to expand this to other ops, otherwise use a libcall.
89 Custom // Use the LowerOperation hook to implement custom lowering.
92 /// This enum indicates whether a types are legal for a target, and if not,
93 /// what action should be used to make them valid.
94 enum LegalizeTypeAction {
95 TypeLegal, // The target natively supports this type.
96 TypePromoteInteger, // Replace this integer with a larger one.
97 TypeExpandInteger, // Split this integer into two of half the size.
98 TypeSoftenFloat, // Convert this float to a same size integer type.
99 TypeExpandFloat, // Split this float into two of half the size.
100 TypeScalarizeVector, // Replace this one-element vector with its element.
101 TypeSplitVector, // Split this vector into two of half the size.
102 TypeWidenVector // This vector should be widened into a larger vector.
105 /// LegalizeKind holds the legalization kind that needs to happen to EVT
106 /// in order to type-legalize it.
107 typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind;
109 /// Enum that describes how the target represents true/false values.
110 enum BooleanContent {
111 UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage.
112 ZeroOrOneBooleanContent, // All bits zero except for bit 0.
113 ZeroOrNegativeOneBooleanContent // All bits equal to bit 0.
116 /// Enum that describes what type of support for selects the target has.
117 enum SelectSupportKind {
118 ScalarValSelect, // The target supports scalar selects (ex: cmov).
119 ScalarCondVectorVal, // The target supports selects with a scalar condition
120 // and vector values (ex: cmov).
121 VectorMaskSelect // The target supports vector selects with a vector
122 // mask (ex: x86 blends).
125 static ISD::NodeType getExtendForContent(BooleanContent Content) {
127 case UndefinedBooleanContent:
128 // Extend by adding rubbish bits.
129 return ISD::ANY_EXTEND;
130 case ZeroOrOneBooleanContent:
131 // Extend by adding zero bits.
132 return ISD::ZERO_EXTEND;
133 case ZeroOrNegativeOneBooleanContent:
134 // Extend by copying the sign bit.
135 return ISD::SIGN_EXTEND;
137 llvm_unreachable("Invalid content kind");
140 /// NOTE: The constructor takes ownership of TLOF.
141 explicit TargetLoweringBase(const TargetMachine &TM,
142 const TargetLoweringObjectFile *TLOF);
143 virtual ~TargetLoweringBase();
146 /// \brief Initialize all of the actions to default values.
150 const TargetMachine &getTargetMachine() const { return TM; }
151 const DataLayout *getDataLayout() const { return DL; }
152 const TargetLoweringObjectFile &getObjFileLowering() const { return TLOF; }
154 bool isBigEndian() const { return !IsLittleEndian; }
155 bool isLittleEndian() const { return IsLittleEndian; }
157 /// Return the pointer type for the given address space, defaults to
158 /// the pointer type from the data layout.
159 /// FIXME: The default needs to be removed once all the code is updated.
160 virtual MVT getPointerTy(uint32_t /*AS*/ = 0) const;
161 unsigned getPointerSizeInBits(uint32_t AS = 0) const;
162 unsigned getPointerTypeSizeInBits(Type *Ty) const;
163 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const;
165 EVT getShiftAmountTy(EVT LHSTy) const;
167 /// Returns the type to be used for the index operand of:
168 /// ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT,
169 /// ISD::INSERT_SUBVECTOR, and ISD::EXTRACT_SUBVECTOR
170 virtual MVT getVectorIdxTy() const {
171 return getPointerTy();
174 /// Return true if the select operation is expensive for this target.
175 bool isSelectExpensive() const { return SelectIsExpensive; }
177 virtual bool isSelectSupported(SelectSupportKind /*kind*/) const {
181 /// Return true if multiple condition registers are available.
182 bool hasMultipleConditionRegisters() const {
183 return HasMultipleConditionRegisters;
186 /// Return true if the target has BitExtract instructions.
187 bool hasExtractBitsInsn() const { return HasExtractBitsInsn; }
189 /// Return true if a vector of the given type should be split
190 /// (TypeSplitVector) instead of promoted (TypePromoteInteger) during type
192 virtual bool shouldSplitVectorType(EVT /*VT*/) const { return false; }
194 // There are two general methods for expanding a BUILD_VECTOR node:
195 // 1. Use SCALAR_TO_VECTOR on the defined scalar values and then shuffle
197 // 2. Build the vector on the stack and then load it.
198 // If this function returns true, then method (1) will be used, subject to
199 // the constraint that all of the necessary shuffles are legal (as determined
200 // by isShuffleMaskLegal). If this function returns false, then method (2) is
201 // always used. The vector type, and the number of defined values, are
204 shouldExpandBuildVectorWithShuffles(EVT /* VT */,
205 unsigned DefinedValues) const {
206 return DefinedValues < 3;
209 /// Return true if integer divide is usually cheaper than a sequence of
210 /// several shifts, adds, and multiplies for this target.
211 bool isIntDivCheap() const { return IntDivIsCheap; }
213 /// Returns true if target has indicated at least one type should be bypassed.
214 bool isSlowDivBypassed() const { return !BypassSlowDivWidths.empty(); }
216 /// Returns map of slow types for division or remainder with corresponding
218 const DenseMap<unsigned int, unsigned int> &getBypassSlowDivWidths() const {
219 return BypassSlowDivWidths;
222 /// Return true if pow2 div is cheaper than a chain of srl/add/sra.
223 bool isPow2DivCheap() const { return Pow2DivIsCheap; }
225 /// Return true if Flow Control is an expensive operation that should be
227 bool isJumpExpensive() const { return JumpIsExpensive; }
229 /// Return true if selects are only cheaper than branches if the branch is
230 /// unlikely to be predicted right.
231 bool isPredictableSelectExpensive() const {
232 return PredictableSelectIsExpensive;
235 /// isLoadBitCastBeneficial() - Return true if the following transform
237 /// fold (conv (load x)) -> (load (conv*)x)
238 /// On architectures that don't natively support some vector loads efficiently,
239 /// casting the load to a smaller vector of larger types and loading
240 /// is more efficient, however, this can be undone by optimizations in
242 virtual bool isLoadBitCastBeneficial(EVT /* Load */, EVT /* Bitcast */) const {
246 /// \brief Return if the target supports combining a
249 /// %andResult = and %val1, #imm-with-one-bit-set;
250 /// %icmpResult = icmp %andResult, 0
251 /// br i1 %icmpResult, label %dest1, label %dest2
253 /// into a single machine instruction of a form like:
255 /// brOnBitSet %register, #bitNumber, dest
257 bool isMaskAndBranchFoldingLegal() const {
258 return MaskAndBranchFoldingIsLegal;
261 /// Return the ValueType of the result of SETCC operations. Also used to
262 /// obtain the target's preferred type for the condition operand of SELECT and
263 /// BRCOND nodes. In the case of BRCOND the argument passed is MVT::Other
264 /// since there are no other operands to get a type hint from.
265 virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
267 /// Return the ValueType for comparison libcalls. Comparions libcalls include
268 /// floating point comparion calls, and Ordered/Unordered check calls on
269 /// floating point numbers.
271 MVT::SimpleValueType getCmpLibcallReturnType() const;
273 /// For targets without i1 registers, this gives the nature of the high-bits
274 /// of boolean values held in types wider than i1.
276 /// "Boolean values" are special true/false values produced by nodes like
277 /// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND.
278 /// Not to be confused with general values promoted from i1. Some cpus
279 /// distinguish between vectors of boolean and scalars; the isVec parameter
280 /// selects between the two kinds. For example on X86 a scalar boolean should
281 /// be zero extended from i1, while the elements of a vector of booleans
282 /// should be sign extended from i1.
283 BooleanContent getBooleanContents(bool isVec) const {
284 return isVec ? BooleanVectorContents : BooleanContents;
287 /// Return target scheduling preference.
288 Sched::Preference getSchedulingPreference() const {
289 return SchedPreferenceInfo;
292 /// Some scheduler, e.g. hybrid, can switch to different scheduling heuristics
293 /// for different nodes. This function returns the preference (or none) for
295 virtual Sched::Preference getSchedulingPreference(SDNode *) const {
299 /// Return the register class that should be used for the specified value
301 virtual const TargetRegisterClass *getRegClassFor(MVT VT) const {
302 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
303 assert(RC && "This value type is not natively supported!");
307 /// Return the 'representative' register class for the specified value
310 /// The 'representative' register class is the largest legal super-reg
311 /// register class for the register class of the value type. For example, on
312 /// i386 the rep register class for i8, i16, and i32 are GR32; while the rep
313 /// register class is GR64 on x86_64.
314 virtual const TargetRegisterClass *getRepRegClassFor(MVT VT) const {
315 const TargetRegisterClass *RC = RepRegClassForVT[VT.SimpleTy];
319 /// Return the cost of the 'representative' register class for the specified
321 virtual uint8_t getRepRegClassCostFor(MVT VT) const {
322 return RepRegClassCostForVT[VT.SimpleTy];
325 /// Return true if the target has native support for the specified value type.
326 /// This means that it has a register that directly holds it without
327 /// promotions or expansions.
328 bool isTypeLegal(EVT VT) const {
329 assert(!VT.isSimple() ||
330 (unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT));
331 return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != nullptr;
334 class ValueTypeActionImpl {
335 /// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum
336 /// that indicates how instruction selection should deal with the type.
337 uint8_t ValueTypeActions[MVT::LAST_VALUETYPE];
340 ValueTypeActionImpl() {
341 std::fill(std::begin(ValueTypeActions), std::end(ValueTypeActions), 0);
344 LegalizeTypeAction getTypeAction(MVT VT) const {
345 return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy];
348 void setTypeAction(MVT VT, LegalizeTypeAction Action) {
349 unsigned I = VT.SimpleTy;
350 ValueTypeActions[I] = Action;
354 const ValueTypeActionImpl &getValueTypeActions() const {
355 return ValueTypeActions;
358 /// Return how we should legalize values of this type, either it is already
359 /// legal (return 'Legal') or we need to promote it to a larger type (return
360 /// 'Promote'), or we need to expand it into multiple registers of smaller
361 /// integer type (return 'Expand'). 'Custom' is not an option.
362 LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const {
363 return getTypeConversion(Context, VT).first;
365 LegalizeTypeAction getTypeAction(MVT VT) const {
366 return ValueTypeActions.getTypeAction(VT);
369 /// For types supported by the target, this is an identity function. For
370 /// types that must be promoted to larger types, this returns the larger type
371 /// to promote to. For integer types that are larger than the largest integer
372 /// register, this contains one step in the expansion to get to the smaller
373 /// register. For illegal floating point types, this returns the integer type
375 EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const {
376 return getTypeConversion(Context, VT).second;
379 /// For types supported by the target, this is an identity function. For
380 /// types that must be expanded (i.e. integer types that are larger than the
381 /// largest integer register or illegal floating point types), this returns
382 /// the largest legal type it will be expanded to.
383 EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const {
384 assert(!VT.isVector());
386 switch (getTypeAction(Context, VT)) {
389 case TypeExpandInteger:
390 VT = getTypeToTransformTo(Context, VT);
393 llvm_unreachable("Type is not legal nor is it to be expanded!");
398 /// Vector types are broken down into some number of legal first class types.
399 /// For example, EVT::v8f32 maps to 2 EVT::v4f32 with Altivec or SSE1, or 8
400 /// promoted EVT::f64 values with the X86 FP stack. Similarly, EVT::v2i64
401 /// turns into 4 EVT::i32 values with both PPC and X86.
403 /// This method returns the number of registers needed, and the VT for each
404 /// register. It also returns the VT and quantity of the intermediate values
405 /// before they are promoted/expanded.
406 unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
408 unsigned &NumIntermediates,
409 MVT &RegisterVT) const;
411 struct IntrinsicInfo {
412 unsigned opc; // target opcode
413 EVT memVT; // memory VT
414 const Value* ptrVal; // value representing memory location
415 int offset; // offset off of ptrVal
416 unsigned align; // alignment
417 bool vol; // is volatile?
418 bool readMem; // reads memory?
419 bool writeMem; // writes memory?
422 /// Given an intrinsic, checks if on the target the intrinsic will need to map
423 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns
424 /// true and store the intrinsic information into the IntrinsicInfo that was
425 /// passed to the function.
426 virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &,
427 unsigned /*Intrinsic*/) const {
431 /// Returns true if the target can instruction select the specified FP
432 /// immediate natively. If false, the legalizer will materialize the FP
433 /// immediate as a load from a constant pool.
434 virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const {
438 /// Targets can use this to indicate that they only support *some*
439 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
440 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to be
442 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/,
447 /// Returns true if the operation can trap for the value type.
449 /// VT must be a legal type. By default, we optimistically assume most
450 /// operations don't trap except for divide and remainder.
451 virtual bool canOpTrap(unsigned Op, EVT VT) const;
453 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
454 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to replace
455 /// a VAND with a constant pool entry.
456 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/,
461 /// Return how this operation should be treated: either it is legal, needs to
462 /// be promoted to a larger size, needs to be expanded to some other code
463 /// sequence, or the target has a custom expander for it.
464 LegalizeAction getOperationAction(unsigned Op, EVT VT) const {
465 if (VT.isExtended()) return Expand;
466 // If a target-specific SDNode requires legalization, require the target
467 // to provide custom legalization for it.
468 if (Op > array_lengthof(OpActions[0])) return Custom;
469 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy;
470 return (LegalizeAction)OpActions[I][Op];
473 /// Return true if the specified operation is legal on this target or can be
474 /// made legal with custom lowering. This is used to help guide high-level
475 /// lowering decisions.
476 bool isOperationLegalOrCustom(unsigned Op, EVT VT) const {
477 return (VT == MVT::Other || isTypeLegal(VT)) &&
478 (getOperationAction(Op, VT) == Legal ||
479 getOperationAction(Op, VT) == Custom);
482 /// Return true if the specified operation is legal on this target or can be
483 /// made legal using promotion. This is used to help guide high-level lowering
485 bool isOperationLegalOrPromote(unsigned Op, EVT VT) const {
486 return (VT == MVT::Other || isTypeLegal(VT)) &&
487 (getOperationAction(Op, VT) == Legal ||
488 getOperationAction(Op, VT) == Promote);
491 /// Return true if the specified operation is illegal on this target or
492 /// unlikely to be made legal with custom lowering. This is used to help guide
493 /// high-level lowering decisions.
494 bool isOperationExpand(unsigned Op, EVT VT) const {
495 return (!isTypeLegal(VT) || getOperationAction(Op, VT) == Expand);
498 /// Return true if the specified operation is legal on this target.
499 bool isOperationLegal(unsigned Op, EVT VT) const {
500 return (VT == MVT::Other || isTypeLegal(VT)) &&
501 getOperationAction(Op, VT) == Legal;
504 /// Return how this load with extension should be treated: either it is legal,
505 /// needs to be promoted to a larger size, needs to be expanded to some other
506 /// code sequence, or the target has a custom expander for it.
507 LegalizeAction getLoadExtAction(unsigned ExtType, MVT VT) const {
508 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE &&
509 "Table isn't big enough!");
510 return (LegalizeAction)LoadExtActions[VT.SimpleTy][ExtType];
513 /// Return true if the specified load with extension is legal on this target.
514 bool isLoadExtLegal(unsigned ExtType, EVT VT) const {
515 return VT.isSimple() &&
516 getLoadExtAction(ExtType, VT.getSimpleVT()) == Legal;
519 /// Return how this store with truncation should be treated: either it is
520 /// legal, needs to be promoted to a larger size, needs to be expanded to some
521 /// other code sequence, or the target has a custom expander for it.
522 LegalizeAction getTruncStoreAction(MVT ValVT, MVT MemVT) const {
523 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE &&
524 "Table isn't big enough!");
525 return (LegalizeAction)TruncStoreActions[ValVT.SimpleTy]
529 /// Return true if the specified store with truncation is legal on this
531 bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const {
532 return isTypeLegal(ValVT) && MemVT.isSimple() &&
533 getTruncStoreAction(ValVT.getSimpleVT(), MemVT.getSimpleVT()) == Legal;
536 /// Return how the indexed load should be treated: either it is legal, needs
537 /// to be promoted to a larger size, needs to be expanded to some other code
538 /// sequence, or the target has a custom expander for it.
540 getIndexedLoadAction(unsigned IdxMode, MVT VT) const {
541 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE &&
542 "Table isn't big enough!");
543 unsigned Ty = (unsigned)VT.SimpleTy;
544 return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4);
547 /// Return true if the specified indexed load is legal on this target.
548 bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const {
549 return VT.isSimple() &&
550 (getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Legal ||
551 getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Custom);
554 /// Return how the indexed store should be treated: either it is legal, needs
555 /// to be promoted to a larger size, needs to be expanded to some other code
556 /// sequence, or the target has a custom expander for it.
558 getIndexedStoreAction(unsigned IdxMode, MVT VT) const {
559 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE &&
560 "Table isn't big enough!");
561 unsigned Ty = (unsigned)VT.SimpleTy;
562 return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f);
565 /// Return true if the specified indexed load is legal on this target.
566 bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const {
567 return VT.isSimple() &&
568 (getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Legal ||
569 getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Custom);
572 /// Return how the condition code should be treated: either it is legal, needs
573 /// to be expanded to some other code sequence, or the target has a custom
576 getCondCodeAction(ISD::CondCode CC, MVT VT) const {
577 assert((unsigned)CC < array_lengthof(CondCodeActions) &&
578 ((unsigned)VT.SimpleTy >> 4) < array_lengthof(CondCodeActions[0]) &&
579 "Table isn't big enough!");
580 // See setCondCodeAction for how this is encoded.
581 uint32_t Shift = 2 * (VT.SimpleTy & 0xF);
582 uint32_t Value = CondCodeActions[CC][VT.SimpleTy >> 4];
583 LegalizeAction Action = (LegalizeAction) ((Value >> Shift) & 0x3);
584 assert(Action != Promote && "Can't promote condition code!");
588 /// Return true if the specified condition code is legal on this target.
589 bool isCondCodeLegal(ISD::CondCode CC, MVT VT) const {
591 getCondCodeAction(CC, VT) == Legal ||
592 getCondCodeAction(CC, VT) == Custom;
596 /// If the action for this operation is to promote, this method returns the
597 /// ValueType to promote to.
598 MVT getTypeToPromoteTo(unsigned Op, MVT VT) const {
599 assert(getOperationAction(Op, VT) == Promote &&
600 "This operation isn't promoted!");
602 // See if this has an explicit type specified.
603 std::map<std::pair<unsigned, MVT::SimpleValueType>,
604 MVT::SimpleValueType>::const_iterator PTTI =
605 PromoteToType.find(std::make_pair(Op, VT.SimpleTy));
606 if (PTTI != PromoteToType.end()) return PTTI->second;
608 assert((VT.isInteger() || VT.isFloatingPoint()) &&
609 "Cannot autopromote this type, add it with AddPromotedToType.");
613 NVT = (MVT::SimpleValueType)(NVT.SimpleTy+1);
614 assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid &&
615 "Didn't find type to promote to!");
616 } while (!isTypeLegal(NVT) ||
617 getOperationAction(Op, NVT) == Promote);
621 /// Return the EVT corresponding to this LLVM type. This is fixed by the LLVM
622 /// operations except for the pointer size. If AllowUnknown is true, this
623 /// will return MVT::Other for types with no EVT counterpart (e.g. structs),
624 /// otherwise it will assert.
625 EVT getValueType(Type *Ty, bool AllowUnknown = false) const {
626 // Lower scalar pointers to native pointer types.
627 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
628 return getPointerTy(PTy->getAddressSpace());
630 if (Ty->isVectorTy()) {
631 VectorType *VTy = cast<VectorType>(Ty);
632 Type *Elm = VTy->getElementType();
633 // Lower vectors of pointers to native pointer types.
634 if (PointerType *PT = dyn_cast<PointerType>(Elm)) {
635 EVT PointerTy(getPointerTy(PT->getAddressSpace()));
636 Elm = PointerTy.getTypeForEVT(Ty->getContext());
639 return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false),
640 VTy->getNumElements());
642 return EVT::getEVT(Ty, AllowUnknown);
645 /// Return the MVT corresponding to this LLVM type. See getValueType.
646 MVT getSimpleValueType(Type *Ty, bool AllowUnknown = false) const {
647 return getValueType(Ty, AllowUnknown).getSimpleVT();
650 /// Return the desired alignment for ByVal or InAlloca aggregate function
651 /// arguments in the caller parameter area. This is the actual alignment, not
653 virtual unsigned getByValTypeAlignment(Type *Ty) const;
655 /// Return the type of registers that this ValueType will eventually require.
656 MVT getRegisterType(MVT VT) const {
657 assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT));
658 return RegisterTypeForVT[VT.SimpleTy];
661 /// Return the type of registers that this ValueType will eventually require.
662 MVT getRegisterType(LLVMContext &Context, EVT VT) const {
664 assert((unsigned)VT.getSimpleVT().SimpleTy <
665 array_lengthof(RegisterTypeForVT));
666 return RegisterTypeForVT[VT.getSimpleVT().SimpleTy];
671 unsigned NumIntermediates;
672 (void)getVectorTypeBreakdown(Context, VT, VT1,
673 NumIntermediates, RegisterVT);
676 if (VT.isInteger()) {
677 return getRegisterType(Context, getTypeToTransformTo(Context, VT));
679 llvm_unreachable("Unsupported extended type!");
682 /// Return the number of registers that this ValueType will eventually
685 /// This is one for any types promoted to live in larger registers, but may be
686 /// more than one for types (like i64) that are split into pieces. For types
687 /// like i140, which are first promoted then expanded, it is the number of
688 /// registers needed to hold all the bits of the original type. For an i140
689 /// on a 32 bit machine this means 5 registers.
690 unsigned getNumRegisters(LLVMContext &Context, EVT VT) const {
692 assert((unsigned)VT.getSimpleVT().SimpleTy <
693 array_lengthof(NumRegistersForVT));
694 return NumRegistersForVT[VT.getSimpleVT().SimpleTy];
699 unsigned NumIntermediates;
700 return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2);
702 if (VT.isInteger()) {
703 unsigned BitWidth = VT.getSizeInBits();
704 unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits();
705 return (BitWidth + RegWidth - 1) / RegWidth;
707 llvm_unreachable("Unsupported extended type!");
710 /// If true, then instruction selection should seek to shrink the FP constant
711 /// of the specified type to a smaller type in order to save space and / or
713 virtual bool ShouldShrinkFPConstant(EVT) const { return true; }
715 /// If true, the target has custom DAG combine transformations that it can
716 /// perform for the specified node.
717 bool hasTargetDAGCombine(ISD::NodeType NT) const {
718 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
719 return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7));
722 /// \brief Get maximum # of store operations permitted for llvm.memset
724 /// This function returns the maximum number of store operations permitted
725 /// to replace a call to llvm.memset. The value is set by the target at the
726 /// performance threshold for such a replacement. If OptSize is true,
727 /// return the limit for functions that have OptSize attribute.
728 unsigned getMaxStoresPerMemset(bool OptSize) const {
729 return OptSize ? MaxStoresPerMemsetOptSize : MaxStoresPerMemset;
732 /// \brief Get maximum # of store operations permitted for llvm.memcpy
734 /// This function returns the maximum number of store operations permitted
735 /// to replace a call to llvm.memcpy. The value is set by the target at the
736 /// performance threshold for such a replacement. If OptSize is true,
737 /// return the limit for functions that have OptSize attribute.
738 unsigned getMaxStoresPerMemcpy(bool OptSize) const {
739 return OptSize ? MaxStoresPerMemcpyOptSize : MaxStoresPerMemcpy;
742 /// \brief Get maximum # of store operations permitted for llvm.memmove
744 /// This function returns the maximum number of store operations permitted
745 /// to replace a call to llvm.memmove. The value is set by the target at the
746 /// performance threshold for such a replacement. If OptSize is true,
747 /// return the limit for functions that have OptSize attribute.
748 unsigned getMaxStoresPerMemmove(bool OptSize) const {
749 return OptSize ? MaxStoresPerMemmoveOptSize : MaxStoresPerMemmove;
752 /// \brief Determine if the target supports unaligned memory accesses.
754 /// This function returns true if the target allows unaligned memory accesses
755 /// of the specified type in the given address space. If true, it also returns
756 /// whether the unaligned memory access is "fast" in the third argument by
757 /// reference. This is used, for example, in situations where an array
758 /// copy/move/set is converted to a sequence of store operations. Its use
759 /// helps to ensure that such replacements don't generate code that causes an
760 /// alignment error (trap) on the target machine.
761 virtual bool allowsUnalignedMemoryAccesses(EVT,
762 unsigned AddrSpace = 0,
763 bool * /*Fast*/ = nullptr) const {
767 /// Returns the target specific optimal type for load and store operations as
768 /// a result of memset, memcpy, and memmove lowering.
770 /// If DstAlign is zero that means it's safe to destination alignment can
771 /// satisfy any constraint. Similarly if SrcAlign is zero it means there isn't
772 /// a need to check it against alignment requirement, probably because the
773 /// source does not need to be loaded. If 'IsMemset' is true, that means it's
774 /// expanding a memset. If 'ZeroMemset' is true, that means it's a memset of
775 /// zero. 'MemcpyStrSrc' indicates whether the memcpy source is constant so it
776 /// does not need to be loaded. It returns EVT::Other if the type should be
777 /// determined using generic target-independent logic.
778 virtual EVT getOptimalMemOpType(uint64_t /*Size*/,
779 unsigned /*DstAlign*/, unsigned /*SrcAlign*/,
782 bool /*MemcpyStrSrc*/,
783 MachineFunction &/*MF*/) const {
787 /// Returns true if it's safe to use load / store of the specified type to
788 /// expand memcpy / memset inline.
790 /// This is mostly true for all types except for some special cases. For
791 /// example, on X86 targets without SSE2 f64 load / store are done with fldl /
792 /// fstpl which also does type conversion. Note the specified type doesn't
793 /// have to be legal as the hook is used before type legalization.
794 virtual bool isSafeMemOpType(MVT /*VT*/) const { return true; }
796 /// Determine if we should use _setjmp or setjmp to implement llvm.setjmp.
797 bool usesUnderscoreSetJmp() const {
798 return UseUnderscoreSetJmp;
801 /// Determine if we should use _longjmp or longjmp to implement llvm.longjmp.
802 bool usesUnderscoreLongJmp() const {
803 return UseUnderscoreLongJmp;
806 /// Return whether the target can generate code for jump tables.
807 bool supportJumpTables() const {
808 return SupportJumpTables;
811 /// Return integer threshold on number of blocks to use jump tables rather
812 /// than if sequence.
813 int getMinimumJumpTableEntries() const {
814 return MinimumJumpTableEntries;
817 /// If a physical register, this specifies the register that
818 /// llvm.savestack/llvm.restorestack should save and restore.
819 unsigned getStackPointerRegisterToSaveRestore() const {
820 return StackPointerRegisterToSaveRestore;
823 /// If a physical register, this returns the register that receives the
824 /// exception address on entry to a landing pad.
825 unsigned getExceptionPointerRegister() const {
826 return ExceptionPointerRegister;
829 /// If a physical register, this returns the register that receives the
830 /// exception typeid on entry to a landing pad.
831 unsigned getExceptionSelectorRegister() const {
832 return ExceptionSelectorRegister;
835 /// Returns the target's jmp_buf size in bytes (if never set, the default is
837 unsigned getJumpBufSize() const {
841 /// Returns the target's jmp_buf alignment in bytes (if never set, the default
843 unsigned getJumpBufAlignment() const {
844 return JumpBufAlignment;
847 /// Return the minimum stack alignment of an argument.
848 unsigned getMinStackArgumentAlignment() const {
849 return MinStackArgumentAlignment;
852 /// Return the minimum function alignment.
853 unsigned getMinFunctionAlignment() const {
854 return MinFunctionAlignment;
857 /// Return the preferred function alignment.
858 unsigned getPrefFunctionAlignment() const {
859 return PrefFunctionAlignment;
862 /// Return the preferred loop alignment.
863 unsigned getPrefLoopAlignment() const {
864 return PrefLoopAlignment;
867 /// Return whether the DAG builder should automatically insert fences and
868 /// reduce ordering for atomics.
869 bool getInsertFencesForAtomic() const {
870 return InsertFencesForAtomic;
873 /// Return true if the target stores stack protector cookies at a fixed offset
874 /// in some non-standard address space, and populates the address space and
875 /// offset as appropriate.
876 virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/,
877 unsigned &/*Offset*/) const {
881 /// Returns the maximal possible offset which can be used for loads / stores
883 virtual unsigned getMaximalGlobalOffset() const {
887 /// Returns the alignment required by global merge on external symbols.
888 /// By default, returns the natural alignment of merged data structure.
889 virtual unsigned getGlobalMergeAlignment(StructType *MergedTy) const {
890 return getDataLayout()->getABITypeAlignment(MergedTy);
893 /// Returns true if a cast between SrcAS and DestAS is a noop.
894 virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const {
898 //===--------------------------------------------------------------------===//
899 /// \name Helpers for TargetTransformInfo implementations
902 /// Get the ISD node that corresponds to the Instruction class opcode.
903 int InstructionOpcodeToISD(unsigned Opcode) const;
905 /// Estimate the cost of type-legalization and the legalized type.
906 std::pair<unsigned, MVT> getTypeLegalizationCost(Type *Ty) const;
910 //===--------------------------------------------------------------------===//
911 /// \name Helpers for load-linked/store-conditional atomic expansion.
914 /// Perform a load-linked operation on Addr, returning a "Value *" with the
915 /// corresponding pointee type. This may entail some non-trivial operations to
916 /// truncate or reconstruct types that will be illegal in the backend. See
917 /// ARMISelLowering for an example implementation.
918 virtual Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
919 AtomicOrdering Ord) const {
920 llvm_unreachable("Load linked unimplemented on this target");
923 /// Perform a store-conditional operation to Addr. Return the status of the
924 /// store. This should be 0 if the store succeeded, non-zero otherwise.
925 virtual Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val,
926 Value *Addr, AtomicOrdering Ord) const {
927 llvm_unreachable("Store conditional unimplemented on this target");
930 /// Return true if the given (atomic) instruction should be expanded by the
931 /// IR-level AtomicExpandLoadLinked pass into a loop involving
932 /// load-linked/store-conditional pairs. Atomic stores will be expanded in the
933 /// same way as "atomic xchg" operations which ignore their output if needed.
934 virtual bool shouldExpandAtomicInIR(Instruction *Inst) const {
939 //===--------------------------------------------------------------------===//
940 // TargetLowering Configuration Methods - These methods should be invoked by
941 // the derived class constructor to configure this object for the target.
944 /// \brief Reset the operation actions based on target options.
945 virtual void resetOperationActions() {}
948 /// Specify how the target extends the result of a boolean value from i1 to a
949 /// wider type. See getBooleanContents.
950 void setBooleanContents(BooleanContent Ty) { BooleanContents = Ty; }
952 /// Specify how the target extends the result of a vector boolean value from a
953 /// vector of i1 to a wider type. See getBooleanContents.
954 void setBooleanVectorContents(BooleanContent Ty) {
955 BooleanVectorContents = Ty;
958 /// Specify the target scheduling preference.
959 void setSchedulingPreference(Sched::Preference Pref) {
960 SchedPreferenceInfo = Pref;
963 /// Indicate whether this target prefers to use _setjmp to implement
964 /// llvm.setjmp or the version without _. Defaults to false.
965 void setUseUnderscoreSetJmp(bool Val) {
966 UseUnderscoreSetJmp = Val;
969 /// Indicate whether this target prefers to use _longjmp to implement
970 /// llvm.longjmp or the version without _. Defaults to false.
971 void setUseUnderscoreLongJmp(bool Val) {
972 UseUnderscoreLongJmp = Val;
975 /// Indicate whether the target can generate code for jump tables.
976 void setSupportJumpTables(bool Val) {
977 SupportJumpTables = Val;
980 /// Indicate the number of blocks to generate jump tables rather than if
982 void setMinimumJumpTableEntries(int Val) {
983 MinimumJumpTableEntries = Val;
986 /// If set to a physical register, this specifies the register that
987 /// llvm.savestack/llvm.restorestack should save and restore.
988 void setStackPointerRegisterToSaveRestore(unsigned R) {
989 StackPointerRegisterToSaveRestore = R;
992 /// If set to a physical register, this sets the register that receives the
993 /// exception address on entry to a landing pad.
994 void setExceptionPointerRegister(unsigned R) {
995 ExceptionPointerRegister = R;
998 /// If set to a physical register, this sets the register that receives the
999 /// exception typeid on entry to a landing pad.
1000 void setExceptionSelectorRegister(unsigned R) {
1001 ExceptionSelectorRegister = R;
1004 /// Tells the code generator not to expand operations into sequences that use
1005 /// the select operations if possible.
1006 void setSelectIsExpensive(bool isExpensive = true) {
1007 SelectIsExpensive = isExpensive;
1010 /// Tells the code generator that the target has multiple (allocatable)
1011 /// condition registers that can be used to store the results of comparisons
1012 /// for use by selects and conditional branches. With multiple condition
1013 /// registers, the code generator will not aggressively sink comparisons into
1014 /// the blocks of their users.
1015 void setHasMultipleConditionRegisters(bool hasManyRegs = true) {
1016 HasMultipleConditionRegisters = hasManyRegs;
1019 /// Tells the code generator that the target has BitExtract instructions.
1020 /// The code generator will aggressively sink "shift"s into the blocks of
1021 /// their users if the users will generate "and" instructions which can be
1022 /// combined with "shift" to BitExtract instructions.
1023 void setHasExtractBitsInsn(bool hasExtractInsn = true) {
1024 HasExtractBitsInsn = hasExtractInsn;
1027 /// Tells the code generator not to expand sequence of operations into a
1028 /// separate sequences that increases the amount of flow control.
1029 void setJumpIsExpensive(bool isExpensive = true) {
1030 JumpIsExpensive = isExpensive;
1033 /// Tells the code generator that integer divide is expensive, and if
1034 /// possible, should be replaced by an alternate sequence of instructions not
1035 /// containing an integer divide.
1036 void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; }
1038 /// Tells the code generator which bitwidths to bypass.
1039 void addBypassSlowDiv(unsigned int SlowBitWidth, unsigned int FastBitWidth) {
1040 BypassSlowDivWidths[SlowBitWidth] = FastBitWidth;
1043 /// Tells the code generator that it shouldn't generate srl/add/sra for a
1044 /// signed divide by power of two, and let the target handle it.
1045 void setPow2DivIsCheap(bool isCheap = true) { Pow2DivIsCheap = isCheap; }
1047 /// Add the specified register class as an available regclass for the
1048 /// specified value type. This indicates the selector can handle values of
1049 /// that class natively.
1050 void addRegisterClass(MVT VT, const TargetRegisterClass *RC) {
1051 assert((unsigned)VT.SimpleTy < array_lengthof(RegClassForVT));
1052 AvailableRegClasses.push_back(std::make_pair(VT, RC));
1053 RegClassForVT[VT.SimpleTy] = RC;
1056 /// Remove all register classes.
1057 void clearRegisterClasses() {
1058 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE * sizeof(TargetRegisterClass*));
1060 AvailableRegClasses.clear();
1063 /// \brief Remove all operation actions.
1064 void clearOperationActions() {
1067 /// Return the largest legal super-reg register class of the register class
1068 /// for the specified type and its associated "cost".
1069 virtual std::pair<const TargetRegisterClass*, uint8_t>
1070 findRepresentativeClass(MVT VT) const;
1072 /// Once all of the register classes are added, this allows us to compute
1073 /// derived properties we expose.
1074 void computeRegisterProperties();
1076 /// Indicate that the specified operation does not work with the specified
1077 /// type and indicate what to do about it.
1078 void setOperationAction(unsigned Op, MVT VT,
1079 LegalizeAction Action) {
1080 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!");
1081 OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action;
1084 /// Indicate that the specified load with extension does not work with the
1085 /// specified type and indicate what to do about it.
1086 void setLoadExtAction(unsigned ExtType, MVT VT,
1087 LegalizeAction Action) {
1088 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE &&
1089 "Table isn't big enough!");
1090 LoadExtActions[VT.SimpleTy][ExtType] = (uint8_t)Action;
1093 /// Indicate that the specified truncating store does not work with the
1094 /// specified type and indicate what to do about it.
1095 void setTruncStoreAction(MVT ValVT, MVT MemVT,
1096 LegalizeAction Action) {
1097 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE &&
1098 "Table isn't big enough!");
1099 TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action;
1102 /// Indicate that the specified indexed load does or does not work with the
1103 /// specified type and indicate what to do abort it.
1105 /// NOTE: All indexed mode loads are initialized to Expand in
1106 /// TargetLowering.cpp
1107 void setIndexedLoadAction(unsigned IdxMode, MVT VT,
1108 LegalizeAction Action) {
1109 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE &&
1110 (unsigned)Action < 0xf && "Table isn't big enough!");
1111 // Load action are kept in the upper half.
1112 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0;
1113 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4;
1116 /// Indicate that the specified indexed store does or does not work with the
1117 /// specified type and indicate what to do about it.
1119 /// NOTE: All indexed mode stores are initialized to Expand in
1120 /// TargetLowering.cpp
1121 void setIndexedStoreAction(unsigned IdxMode, MVT VT,
1122 LegalizeAction Action) {
1123 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE &&
1124 (unsigned)Action < 0xf && "Table isn't big enough!");
1125 // Store action are kept in the lower half.
1126 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f;
1127 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action);
1130 /// Indicate that the specified condition code is or isn't supported on the
1131 /// target and indicate what to do about it.
1132 void setCondCodeAction(ISD::CondCode CC, MVT VT,
1133 LegalizeAction Action) {
1134 assert(VT < MVT::LAST_VALUETYPE &&
1135 (unsigned)CC < array_lengthof(CondCodeActions) &&
1136 "Table isn't big enough!");
1137 /// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 32-bit
1138 /// value and the upper 27 bits index into the second dimension of the array
1139 /// to select what 32-bit value to use.
1140 uint32_t Shift = 2 * (VT.SimpleTy & 0xF);
1141 CondCodeActions[CC][VT.SimpleTy >> 4] &= ~((uint32_t)0x3 << Shift);
1142 CondCodeActions[CC][VT.SimpleTy >> 4] |= (uint32_t)Action << Shift;
1145 /// If Opc/OrigVT is specified as being promoted, the promotion code defaults
1146 /// to trying a larger integer/fp until it can find one that works. If that
1147 /// default is insufficient, this method can be used by the target to override
1149 void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) {
1150 PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy;
1153 /// Targets should invoke this method for each target independent node that
1154 /// they want to provide a custom DAG combiner for by implementing the
1155 /// PerformDAGCombine virtual method.
1156 void setTargetDAGCombine(ISD::NodeType NT) {
1157 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
1158 TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7);
1161 /// Set the target's required jmp_buf buffer size (in bytes); default is 200
1162 void setJumpBufSize(unsigned Size) {
1166 /// Set the target's required jmp_buf buffer alignment (in bytes); default is
1168 void setJumpBufAlignment(unsigned Align) {
1169 JumpBufAlignment = Align;
1172 /// Set the target's minimum function alignment (in log2(bytes))
1173 void setMinFunctionAlignment(unsigned Align) {
1174 MinFunctionAlignment = Align;
1177 /// Set the target's preferred function alignment. This should be set if
1178 /// there is a performance benefit to higher-than-minimum alignment (in
1180 void setPrefFunctionAlignment(unsigned Align) {
1181 PrefFunctionAlignment = Align;
1184 /// Set the target's preferred loop alignment. Default alignment is zero, it
1185 /// means the target does not care about loop alignment. The alignment is
1186 /// specified in log2(bytes).
1187 void setPrefLoopAlignment(unsigned Align) {
1188 PrefLoopAlignment = Align;
1191 /// Set the minimum stack alignment of an argument (in log2(bytes)).
1192 void setMinStackArgumentAlignment(unsigned Align) {
1193 MinStackArgumentAlignment = Align;
1196 /// Set if the DAG builder should automatically insert fences and reduce the
1197 /// order of atomic memory operations to Monotonic.
1198 void setInsertFencesForAtomic(bool fence) {
1199 InsertFencesForAtomic = fence;
1203 //===--------------------------------------------------------------------===//
1204 // Addressing mode description hooks (used by LSR etc).
1207 /// CodeGenPrepare sinks address calculations into the same BB as Load/Store
1208 /// instructions reading the address. This allows as much computation as
1209 /// possible to be done in the address mode for that operand. This hook lets
1210 /// targets also pass back when this should be done on intrinsics which
1212 virtual bool GetAddrModeArguments(IntrinsicInst * /*I*/,
1213 SmallVectorImpl<Value*> &/*Ops*/,
1214 Type *&/*AccessTy*/) const {
1218 /// This represents an addressing mode of:
1219 /// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
1220 /// If BaseGV is null, there is no BaseGV.
1221 /// If BaseOffs is zero, there is no base offset.
1222 /// If HasBaseReg is false, there is no base register.
1223 /// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with
1226 GlobalValue *BaseGV;
1230 AddrMode() : BaseGV(nullptr), BaseOffs(0), HasBaseReg(false), Scale(0) {}
1233 /// Return true if the addressing mode represented by AM is legal for this
1234 /// target, for a load/store of the specified type.
1236 /// The type may be VoidTy, in which case only return true if the addressing
1237 /// mode is legal for a load/store of any legal type. TODO: Handle
1238 /// pre/postinc as well.
1239 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const;
1241 /// \brief Return the cost of the scaling factor used in the addressing mode
1242 /// represented by AM for this target, for a load/store of the specified type.
1244 /// If the AM is supported, the return value must be >= 0.
1245 /// If the AM is not supported, it returns a negative value.
1246 /// TODO: Handle pre/postinc as well.
1247 virtual int getScalingFactorCost(const AddrMode &AM, Type *Ty) const {
1248 // Default: assume that any scaling factor used in a legal AM is free.
1249 if (isLegalAddressingMode(AM, Ty)) return 0;
1253 /// Return true if the specified immediate is legal icmp immediate, that is
1254 /// the target has icmp instructions which can compare a register against the
1255 /// immediate without having to materialize the immediate into a register.
1256 virtual bool isLegalICmpImmediate(int64_t) const {
1260 /// Return true if the specified immediate is legal add immediate, that is the
1261 /// target has add instructions which can add a register with the immediate
1262 /// without having to materialize the immediate into a register.
1263 virtual bool isLegalAddImmediate(int64_t) const {
1267 /// Return true if it's significantly cheaper to shift a vector by a uniform
1268 /// scalar than by an amount which will vary across each lane. On x86, for
1269 /// example, there is a "psllw" instruction for the former case, but no simple
1270 /// instruction for a general "a << b" operation on vectors.
1271 virtual bool isVectorShiftByScalarCheap(Type *Ty) const {
1275 /// Return true if it's free to truncate a value of type Ty1 to type
1276 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
1277 /// by referencing its sub-register AX.
1278 virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const {
1282 /// Return true if a truncation from Ty1 to Ty2 is permitted when deciding
1283 /// whether a call is in tail position. Typically this means that both results
1284 /// would be assigned to the same register or stack slot, but it could mean
1285 /// the target performs adequate checks of its own before proceeding with the
1287 virtual bool allowTruncateForTailCall(Type * /*Ty1*/, Type * /*Ty2*/) const {
1291 virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const {
1295 /// Return true if any actual instruction that defines a value of type Ty1
1296 /// implicitly zero-extends the value to Ty2 in the result register.
1298 /// This does not necessarily include registers defined in unknown ways, such
1299 /// as incoming arguments, or copies from unknown virtual registers. Also, if
1300 /// isTruncateFree(Ty2, Ty1) is true, this does not necessarily apply to
1301 /// truncate instructions. e.g. on x86-64, all instructions that define 32-bit
1302 /// values implicit zero-extend the result out to 64 bits.
1303 virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const {
1307 virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const {
1311 /// Return true if the target supplies and combines to a paired load
1312 /// two loaded values of type LoadedType next to each other in memory.
1313 /// RequiredAlignment gives the minimal alignment constraints that must be met
1314 /// to be able to select this paired load.
1316 /// This information is *not* used to generate actual paired loads, but it is
1317 /// used to generate a sequence of loads that is easier to combine into a
1319 /// For instance, something like this:
1320 /// a = load i64* addr
1321 /// b = trunc i64 a to i32
1322 /// c = lshr i64 a, 32
1323 /// d = trunc i64 c to i32
1324 /// will be optimized into:
1325 /// b = load i32* addr1
1326 /// d = load i32* addr2
1327 /// Where addr1 = addr2 +/- sizeof(i32).
1329 /// In other words, unless the target performs a post-isel load combining,
1330 /// this information should not be provided because it will generate more
1332 virtual bool hasPairedLoad(Type * /*LoadedType*/,
1333 unsigned & /*RequiredAligment*/) const {
1337 virtual bool hasPairedLoad(EVT /*LoadedType*/,
1338 unsigned & /*RequiredAligment*/) const {
1342 /// Return true if zero-extending the specific node Val to type VT2 is free
1343 /// (either because it's implicitly zero-extended such as ARM ldrb / ldrh or
1344 /// because it's folded such as X86 zero-extending loads).
1345 virtual bool isZExtFree(SDValue Val, EVT VT2) const {
1346 return isZExtFree(Val.getValueType(), VT2);
1349 /// Return true if an fneg operation is free to the point where it is never
1350 /// worthwhile to replace it with a bitwise operation.
1351 virtual bool isFNegFree(EVT VT) const {
1352 assert(VT.isFloatingPoint());
1356 /// Return true if an fabs operation is free to the point where it is never
1357 /// worthwhile to replace it with a bitwise operation.
1358 virtual bool isFAbsFree(EVT VT) const {
1359 assert(VT.isFloatingPoint());
1363 /// Return true if an FMA operation is faster than a pair of fmul and fadd
1364 /// instructions. fmuladd intrinsics will be expanded to FMAs when this method
1365 /// returns true, otherwise fmuladd is expanded to fmul + fadd.
1367 /// NOTE: This may be called before legalization on types for which FMAs are
1368 /// not legal, but should return true if those types will eventually legalize
1369 /// to types that support FMAs. After legalization, it will only be called on
1370 /// types that support FMAs (via Legal or Custom actions)
1371 virtual bool isFMAFasterThanFMulAndFAdd(EVT) const {
1375 /// Return true if it's profitable to narrow operations of type VT1 to
1376 /// VT2. e.g. on x86, it's profitable to narrow from i32 to i8 but not from
1378 virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const {
1382 /// \brief Return true if it is beneficial to convert a load of a constant to
1383 /// just the constant itself.
1384 /// On some targets it might be more efficient to use a combination of
1385 /// arithmetic instructions to materialize the constant instead of loading it
1386 /// from a constant pool.
1387 virtual bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
1391 //===--------------------------------------------------------------------===//
1392 // Runtime Library hooks
1395 /// Rename the default libcall routine name for the specified libcall.
1396 void setLibcallName(RTLIB::Libcall Call, const char *Name) {
1397 LibcallRoutineNames[Call] = Name;
1400 /// Get the libcall routine name for the specified libcall.
1401 const char *getLibcallName(RTLIB::Libcall Call) const {
1402 return LibcallRoutineNames[Call];
1405 /// Override the default CondCode to be used to test the result of the
1406 /// comparison libcall against zero.
1407 void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) {
1408 CmpLibcallCCs[Call] = CC;
1411 /// Get the CondCode that's to be used to test the result of the comparison
1412 /// libcall against zero.
1413 ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const {
1414 return CmpLibcallCCs[Call];
1417 /// Set the CallingConv that should be used for the specified libcall.
1418 void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) {
1419 LibcallCallingConvs[Call] = CC;
1422 /// Get the CallingConv that should be used for the specified libcall.
1423 CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const {
1424 return LibcallCallingConvs[Call];
1428 const TargetMachine &TM;
1429 const DataLayout *DL;
1430 const TargetLoweringObjectFile &TLOF;
1432 /// True if this is a little endian target.
1433 bool IsLittleEndian;
1435 /// Tells the code generator not to expand operations into sequences that use
1436 /// the select operations if possible.
1437 bool SelectIsExpensive;
1439 /// Tells the code generator that the target has multiple (allocatable)
1440 /// condition registers that can be used to store the results of comparisons
1441 /// for use by selects and conditional branches. With multiple condition
1442 /// registers, the code generator will not aggressively sink comparisons into
1443 /// the blocks of their users.
1444 bool HasMultipleConditionRegisters;
1446 /// Tells the code generator that the target has BitExtract instructions.
1447 /// The code generator will aggressively sink "shift"s into the blocks of
1448 /// their users if the users will generate "and" instructions which can be
1449 /// combined with "shift" to BitExtract instructions.
1450 bool HasExtractBitsInsn;
1452 /// Tells the code generator not to expand integer divides by constants into a
1453 /// sequence of muls, adds, and shifts. This is a hack until a real cost
1454 /// model is in place. If we ever optimize for size, this will be set to true
1455 /// unconditionally.
1458 /// Tells the code generator to bypass slow divide or remainder
1459 /// instructions. For example, BypassSlowDivWidths[32,8] tells the code
1460 /// generator to bypass 32-bit integer div/rem with an 8-bit unsigned integer
1461 /// div/rem when the operands are positive and less than 256.
1462 DenseMap <unsigned int, unsigned int> BypassSlowDivWidths;
1464 /// Tells the code generator that it shouldn't generate srl/add/sra for a
1465 /// signed divide by power of two, and let the target handle it.
1466 bool Pow2DivIsCheap;
1468 /// Tells the code generator that it shouldn't generate extra flow control
1469 /// instructions and should attempt to combine flow control instructions via
1471 bool JumpIsExpensive;
1473 /// This target prefers to use _setjmp to implement llvm.setjmp.
1475 /// Defaults to false.
1476 bool UseUnderscoreSetJmp;
1478 /// This target prefers to use _longjmp to implement llvm.longjmp.
1480 /// Defaults to false.
1481 bool UseUnderscoreLongJmp;
1483 /// Whether the target can generate code for jumptables. If it's not true,
1484 /// then each jumptable must be lowered into if-then-else's.
1485 bool SupportJumpTables;
1487 /// Number of blocks threshold to use jump tables.
1488 int MinimumJumpTableEntries;
1490 /// Information about the contents of the high-bits in boolean values held in
1491 /// a type wider than i1. See getBooleanContents.
1492 BooleanContent BooleanContents;
1494 /// Information about the contents of the high-bits in boolean vector values
1495 /// when the element type is wider than i1. See getBooleanContents.
1496 BooleanContent BooleanVectorContents;
1498 /// The target scheduling preference: shortest possible total cycles or lowest
1500 Sched::Preference SchedPreferenceInfo;
1502 /// The size, in bytes, of the target's jmp_buf buffers
1503 unsigned JumpBufSize;
1505 /// The alignment, in bytes, of the target's jmp_buf buffers
1506 unsigned JumpBufAlignment;
1508 /// The minimum alignment that any argument on the stack needs to have.
1509 unsigned MinStackArgumentAlignment;
1511 /// The minimum function alignment (used when optimizing for size, and to
1512 /// prevent explicitly provided alignment from leading to incorrect code).
1513 unsigned MinFunctionAlignment;
1515 /// The preferred function alignment (used when alignment unspecified and
1516 /// optimizing for speed).
1517 unsigned PrefFunctionAlignment;
1519 /// The preferred loop alignment.
1520 unsigned PrefLoopAlignment;
1522 /// Whether the DAG builder should automatically insert fences and reduce
1523 /// ordering for atomics. (This will be set for for most architectures with
1524 /// weak memory ordering.)
1525 bool InsertFencesForAtomic;
1527 /// If set to a physical register, this specifies the register that
1528 /// llvm.savestack/llvm.restorestack should save and restore.
1529 unsigned StackPointerRegisterToSaveRestore;
1531 /// If set to a physical register, this specifies the register that receives
1532 /// the exception address on entry to a landing pad.
1533 unsigned ExceptionPointerRegister;
1535 /// If set to a physical register, this specifies the register that receives
1536 /// the exception typeid on entry to a landing pad.
1537 unsigned ExceptionSelectorRegister;
1539 /// This indicates the default register class to use for each ValueType the
1540 /// target supports natively.
1541 const TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE];
1542 unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE];
1543 MVT RegisterTypeForVT[MVT::LAST_VALUETYPE];
1545 /// This indicates the "representative" register class to use for each
1546 /// ValueType the target supports natively. This information is used by the
1547 /// scheduler to track register pressure. By default, the representative
1548 /// register class is the largest legal super-reg register class of the
1549 /// register class of the specified type. e.g. On x86, i8, i16, and i32's
1550 /// representative class would be GR32.
1551 const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE];
1553 /// This indicates the "cost" of the "representative" register class for each
1554 /// ValueType. The cost is used by the scheduler to approximate register
1556 uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE];
1558 /// For any value types we are promoting or expanding, this contains the value
1559 /// type that we are changing to. For Expanded types, this contains one step
1560 /// of the expand (e.g. i64 -> i32), even if there are multiple steps required
1561 /// (e.g. i64 -> i16). For types natively supported by the system, this holds
1562 /// the same type (e.g. i32 -> i32).
1563 MVT TransformToType[MVT::LAST_VALUETYPE];
1565 /// For each operation and each value type, keep a LegalizeAction that
1566 /// indicates how instruction selection should deal with the operation. Most
1567 /// operations are Legal (aka, supported natively by the target), but
1568 /// operations that are not should be described. Note that operations on
1569 /// non-legal value types are not described here.
1570 uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END];
1572 /// For each load extension type and each value type, keep a LegalizeAction
1573 /// that indicates how instruction selection should deal with a load of a
1574 /// specific value type and extension type.
1575 uint8_t LoadExtActions[MVT::LAST_VALUETYPE][ISD::LAST_LOADEXT_TYPE];
1577 /// For each value type pair keep a LegalizeAction that indicates whether a
1578 /// truncating store of a specific value type and truncating type is legal.
1579 uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE];
1581 /// For each indexed mode and each value type, keep a pair of LegalizeAction
1582 /// that indicates how instruction selection should deal with the load /
1585 /// The first dimension is the value_type for the reference. The second
1586 /// dimension represents the various modes for load store.
1587 uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE];
1589 /// For each condition code (ISD::CondCode) keep a LegalizeAction that
1590 /// indicates how instruction selection should deal with the condition code.
1592 /// Because each CC action takes up 2 bits, we need to have the array size be
1593 /// large enough to fit all of the value types. This can be done by rounding
1594 /// up the MVT::LAST_VALUETYPE value to the next multiple of 16.
1595 uint32_t CondCodeActions[ISD::SETCC_INVALID][(MVT::LAST_VALUETYPE + 15) / 16];
1597 ValueTypeActionImpl ValueTypeActions;
1601 getTypeConversion(LLVMContext &Context, EVT VT) const {
1602 // If this is a simple type, use the ComputeRegisterProp mechanism.
1603 if (VT.isSimple()) {
1604 MVT SVT = VT.getSimpleVT();
1605 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
1606 MVT NVT = TransformToType[SVT.SimpleTy];
1607 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
1611 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)
1612 && "Promote may not follow Expand or Promote");
1614 if (LA == TypeSplitVector)
1615 return LegalizeKind(LA, EVT::getVectorVT(Context,
1616 SVT.getVectorElementType(),
1617 SVT.getVectorNumElements()/2));
1618 if (LA == TypeScalarizeVector)
1619 return LegalizeKind(LA, SVT.getVectorElementType());
1620 return LegalizeKind(LA, NVT);
1623 // Handle Extended Scalar Types.
1624 if (!VT.isVector()) {
1625 assert(VT.isInteger() && "Float types must be simple");
1626 unsigned BitSize = VT.getSizeInBits();
1627 // First promote to a power-of-two size, then expand if necessary.
1628 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
1629 EVT NVT = VT.getRoundIntegerType(Context);
1630 assert(NVT != VT && "Unable to round integer VT");
1631 LegalizeKind NextStep = getTypeConversion(Context, NVT);
1632 // Avoid multi-step promotion.
1633 if (NextStep.first == TypePromoteInteger) return NextStep;
1634 // Return rounded integer type.
1635 return LegalizeKind(TypePromoteInteger, NVT);
1638 return LegalizeKind(TypeExpandInteger,
1639 EVT::getIntegerVT(Context, VT.getSizeInBits()/2));
1642 // Handle vector types.
1643 unsigned NumElts = VT.getVectorNumElements();
1644 EVT EltVT = VT.getVectorElementType();
1646 // Vectors with only one element are always scalarized.
1648 return LegalizeKind(TypeScalarizeVector, EltVT);
1650 // Try to widen vector elements until the element type is a power of two and
1651 // promote it to a legal type later on, for example:
1652 // <3 x i8> -> <4 x i8> -> <4 x i32>
1653 if (EltVT.isInteger()) {
1654 // Vectors with a number of elements that is not a power of two are always
1655 // widened, for example <3 x i8> -> <4 x i8>.
1656 if (!VT.isPow2VectorType()) {
1657 NumElts = (unsigned)NextPowerOf2(NumElts);
1658 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
1659 return LegalizeKind(TypeWidenVector, NVT);
1662 // Examine the element type.
1663 LegalizeKind LK = getTypeConversion(Context, EltVT);
1665 // If type is to be expanded, split the vector.
1666 // <4 x i140> -> <2 x i140>
1667 if (LK.first == TypeExpandInteger)
1668 return LegalizeKind(TypeSplitVector,
1669 EVT::getVectorVT(Context, EltVT, NumElts / 2));
1671 // Promote the integer element types until a legal vector type is found
1672 // or until the element integer type is too big. If a legal type was not
1673 // found, fallback to the usual mechanism of widening/splitting the
1675 EVT OldEltVT = EltVT;
1677 // Increase the bitwidth of the element to the next pow-of-two
1678 // (which is greater than 8 bits).
1679 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()
1680 ).getRoundIntegerType(Context);
1682 // Stop trying when getting a non-simple element type.
1683 // Note that vector elements may be greater than legal vector element
1684 // types. Example: X86 XMM registers hold 64bit element on 32bit
1686 if (!EltVT.isSimple()) break;
1688 // Build a new vector type and check if it is legal.
1689 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1690 // Found a legal promoted vector type.
1691 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
1692 return LegalizeKind(TypePromoteInteger,
1693 EVT::getVectorVT(Context, EltVT, NumElts));
1696 // Reset the type to the unexpanded type if we did not find a legal vector
1697 // type with a promoted vector element type.
1701 // Try to widen the vector until a legal type is found.
1702 // If there is no wider legal type, split the vector.
1704 // Round up to the next power of 2.
1705 NumElts = (unsigned)NextPowerOf2(NumElts);
1707 // If there is no simple vector type with this many elements then there
1708 // cannot be a larger legal vector type. Note that this assumes that
1709 // there are no skipped intermediate vector types in the simple types.
1710 if (!EltVT.isSimple()) break;
1711 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1712 if (LargerVector == MVT()) break;
1714 // If this type is legal then widen the vector.
1715 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
1716 return LegalizeKind(TypeWidenVector, LargerVector);
1719 // Widen odd vectors to next power of two.
1720 if (!VT.isPow2VectorType()) {
1721 EVT NVT = VT.getPow2VectorType(Context);
1722 return LegalizeKind(TypeWidenVector, NVT);
1725 // Vectors with illegal element types are expanded.
1726 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
1727 return LegalizeKind(TypeSplitVector, NVT);
1731 std::vector<std::pair<MVT, const TargetRegisterClass*> > AvailableRegClasses;
1733 /// Targets can specify ISD nodes that they would like PerformDAGCombine
1734 /// callbacks for by calling setTargetDAGCombine(), which sets a bit in this
1737 TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT];
1739 /// For operations that must be promoted to a specific type, this holds the
1740 /// destination type. This map should be sparse, so don't hold it as an
1743 /// Targets add entries to this map with AddPromotedToType(..), clients access
1744 /// this with getTypeToPromoteTo(..).
1745 std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType>
1748 /// Stores the name each libcall.
1749 const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL];
1751 /// The ISD::CondCode that should be used to test the result of each of the
1752 /// comparison libcall against zero.
1753 ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL];
1755 /// Stores the CallingConv that should be used for each libcall.
1756 CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL];
1759 /// \brief Specify maximum number of store instructions per memset call.
1761 /// When lowering \@llvm.memset this field specifies the maximum number of
1762 /// store operations that may be substituted for the call to memset. Targets
1763 /// must set this value based on the cost threshold for that target. Targets
1764 /// should assume that the memset will be done using as many of the largest
1765 /// store operations first, followed by smaller ones, if necessary, per
1766 /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine
1767 /// with 16-bit alignment would result in four 2-byte stores and one 1-byte
1768 /// store. This only applies to setting a constant array of a constant size.
1769 unsigned MaxStoresPerMemset;
1771 /// Maximum number of stores operations that may be substituted for the call
1772 /// to memset, used for functions with OptSize attribute.
1773 unsigned MaxStoresPerMemsetOptSize;
1775 /// \brief Specify maximum bytes of store instructions per memcpy call.
1777 /// When lowering \@llvm.memcpy this field specifies the maximum number of
1778 /// store operations that may be substituted for a call to memcpy. Targets
1779 /// must set this value based on the cost threshold for that target. Targets
1780 /// should assume that the memcpy will be done using as many of the largest
1781 /// store operations first, followed by smaller ones, if necessary, per
1782 /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine
1783 /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store
1784 /// and one 1-byte store. This only applies to copying a constant array of
1786 unsigned MaxStoresPerMemcpy;
1788 /// Maximum number of store operations that may be substituted for a call to
1789 /// memcpy, used for functions with OptSize attribute.
1790 unsigned MaxStoresPerMemcpyOptSize;
1792 /// \brief Specify maximum bytes of store instructions per memmove call.
1794 /// When lowering \@llvm.memmove this field specifies the maximum number of
1795 /// store instructions that may be substituted for a call to memmove. Targets
1796 /// must set this value based on the cost threshold for that target. Targets
1797 /// should assume that the memmove will be done using as many of the largest
1798 /// store operations first, followed by smaller ones, if necessary, per
1799 /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine
1800 /// with 8-bit alignment would result in nine 1-byte stores. This only
1801 /// applies to copying a constant array of constant size.
1802 unsigned MaxStoresPerMemmove;
1804 /// Maximum number of store instructions that may be substituted for a call to
1805 /// memmove, used for functions with OpSize attribute.
1806 unsigned MaxStoresPerMemmoveOptSize;
1808 /// Tells the code generator that select is more expensive than a branch if
1809 /// the branch is usually predicted right.
1810 bool PredictableSelectIsExpensive;
1812 /// MaskAndBranchFoldingIsLegal - Indicates if the target supports folding
1813 /// a mask of a single bit, a compare, and a branch into a single instruction.
1814 bool MaskAndBranchFoldingIsLegal;
1817 /// Return true if the value types that can be represented by the specified
1818 /// register class are all legal.
1819 bool isLegalRC(const TargetRegisterClass *RC) const;
1821 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
1822 /// sequence of memory operands that is recognized by PrologEpilogInserter.
1823 MachineBasicBlock *emitPatchPoint(MachineInstr *MI, MachineBasicBlock *MBB) const;
1826 /// This class defines information used to lower LLVM code to legal SelectionDAG
1827 /// operators that the target instruction selector can accept natively.
1829 /// This class also defines callbacks that targets must implement to lower
1830 /// target-specific constructs to SelectionDAG operators.
1831 class TargetLowering : public TargetLoweringBase {
1832 TargetLowering(const TargetLowering&) LLVM_DELETED_FUNCTION;
1833 void operator=(const TargetLowering&) LLVM_DELETED_FUNCTION;
1836 /// NOTE: The constructor takes ownership of TLOF.
1837 explicit TargetLowering(const TargetMachine &TM,
1838 const TargetLoweringObjectFile *TLOF);
1840 /// Returns true by value, base pointer and offset pointer and addressing mode
1841 /// by reference if the node's address can be legally represented as
1842 /// pre-indexed load / store address.
1843 virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/,
1844 SDValue &/*Offset*/,
1845 ISD::MemIndexedMode &/*AM*/,
1846 SelectionDAG &/*DAG*/) const {
1850 /// Returns true by value, base pointer and offset pointer and addressing mode
1851 /// by reference if this node can be combined with a load / store to form a
1852 /// post-indexed load / store.
1853 virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/,
1855 SDValue &/*Offset*/,
1856 ISD::MemIndexedMode &/*AM*/,
1857 SelectionDAG &/*DAG*/) const {
1861 /// Return the entry encoding for a jump table in the current function. The
1862 /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum.
1863 virtual unsigned getJumpTableEncoding() const;
1865 virtual const MCExpr *
1866 LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/,
1867 const MachineBasicBlock * /*MBB*/, unsigned /*uid*/,
1868 MCContext &/*Ctx*/) const {
1869 llvm_unreachable("Need to implement this hook if target has custom JTIs");
1872 /// Returns relocation base for the given PIC jumptable.
1873 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
1874 SelectionDAG &DAG) const;
1876 /// This returns the relocation base for the given PIC jumptable, the same as
1877 /// getPICJumpTableRelocBase, but as an MCExpr.
1878 virtual const MCExpr *
1879 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1880 unsigned JTI, MCContext &Ctx) const;
1882 /// Return true if folding a constant offset with the given GlobalAddress is
1883 /// legal. It is frequently not legal in PIC relocation models.
1884 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
1886 bool isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
1887 SDValue &Chain) const;
1889 void softenSetCCOperands(SelectionDAG &DAG, EVT VT,
1890 SDValue &NewLHS, SDValue &NewRHS,
1891 ISD::CondCode &CCCode, SDLoc DL) const;
1893 /// Returns a pair of (return value, chain).
1894 std::pair<SDValue, SDValue> makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC,
1895 EVT RetVT, const SDValue *Ops,
1896 unsigned NumOps, bool isSigned,
1897 SDLoc dl, bool doesNotReturn = false,
1898 bool isReturnValueUsed = true) const;
1900 //===--------------------------------------------------------------------===//
1901 // TargetLowering Optimization Methods
1904 /// A convenience struct that encapsulates a DAG, and two SDValues for
1905 /// returning information from TargetLowering to its clients that want to
1907 struct TargetLoweringOpt {
1914 explicit TargetLoweringOpt(SelectionDAG &InDAG,
1916 DAG(InDAG), LegalTys(LT), LegalOps(LO) {}
1918 bool LegalTypes() const { return LegalTys; }
1919 bool LegalOperations() const { return LegalOps; }
1921 bool CombineTo(SDValue O, SDValue N) {
1927 /// Check to see if the specified operand of the specified instruction is a
1928 /// constant integer. If so, check to see if there are any bits set in the
1929 /// constant that are not demanded. If so, shrink the constant and return
1931 bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded);
1933 /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free. This
1934 /// uses isZExtFree and ZERO_EXTEND for the widening cast, but it could be
1935 /// generalized for targets with other types of implicit widening casts.
1936 bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded,
1940 /// Look at Op. At this point, we know that only the DemandedMask bits of the
1941 /// result of Op are ever used downstream. If we can use this information to
1942 /// simplify Op, create a new simplified DAG node and return true, returning
1943 /// the original and new nodes in Old and New. Otherwise, analyze the
1944 /// expression and return a mask of KnownOne and KnownZero bits for the
1945 /// expression (used to simplify the caller). The KnownZero/One bits may only
1946 /// be accurate for those bits in the DemandedMask.
1947 bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask,
1948 APInt &KnownZero, APInt &KnownOne,
1949 TargetLoweringOpt &TLO, unsigned Depth = 0) const;
1951 /// Determine which of the bits specified in Mask are known to be either zero
1952 /// or one and return them in the KnownZero/KnownOne bitsets.
1953 virtual void computeKnownBitsForTargetNode(const SDValue Op,
1956 const SelectionDAG &DAG,
1957 unsigned Depth = 0) const;
1959 /// This method can be implemented by targets that want to expose additional
1960 /// information about sign bits to the DAG Combiner.
1961 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
1962 const SelectionDAG &DAG,
1963 unsigned Depth = 0) const;
1965 struct DAGCombinerInfo {
1966 void *DC; // The DAG Combiner object.
1968 bool CalledByLegalizer;
1972 DAGCombinerInfo(SelectionDAG &dag, CombineLevel level, bool cl, void *dc)
1973 : DC(dc), Level(level), CalledByLegalizer(cl), DAG(dag) {}
1975 bool isBeforeLegalize() const { return Level == BeforeLegalizeTypes; }
1976 bool isBeforeLegalizeOps() const { return Level < AfterLegalizeVectorOps; }
1977 bool isAfterLegalizeVectorOps() const {
1978 return Level == AfterLegalizeDAG;
1980 CombineLevel getDAGCombineLevel() { return Level; }
1981 bool isCalledByLegalizer() const { return CalledByLegalizer; }
1983 void AddToWorklist(SDNode *N);
1984 void RemoveFromWorklist(SDNode *N);
1985 SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To,
1987 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true);
1988 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true);
1990 void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO);
1993 /// Return if the N is a constant or constant vector equal to the true value
1994 /// from getBooleanContents().
1995 bool isConstTrueVal(const SDNode *N) const;
1997 /// Return if the N is a constant or constant vector equal to the false value
1998 /// from getBooleanContents().
1999 bool isConstFalseVal(const SDNode *N) const;
2001 /// Try to simplify a setcc built with the specified operands and cc. If it is
2002 /// unable to simplify it, return a null SDValue.
2003 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
2004 ISD::CondCode Cond, bool foldBooleans,
2005 DAGCombinerInfo &DCI, SDLoc dl) const;
2007 /// Returns true (and the GlobalValue and the offset) if the node is a
2008 /// GlobalAddress + offset.
2010 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
2012 /// This method will be invoked for all target nodes and for any
2013 /// target-independent nodes that the target has registered with invoke it
2016 /// The semantics are as follows:
2018 /// SDValue.Val == 0 - No change was made
2019 /// SDValue.Val == N - N was replaced, is dead, and is already handled.
2020 /// otherwise - N should be replaced by the returned Operand.
2022 /// In addition, methods provided by DAGCombinerInfo may be used to perform
2023 /// more complex transformations.
2025 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
2027 /// Return true if it is profitable to move a following shift through this
2028 // node, adjusting any immediate operands as necessary to preserve semantics.
2029 // This transformation may not be desirable if it disrupts a particularly
2030 // auspicious target-specific tree (e.g. bitfield extraction in AArch64).
2031 // By default, it returns true.
2032 virtual bool isDesirableToCommuteWithShift(const SDNode *N /*Op*/) const {
2036 /// Return true if the target has native support for the specified value type
2037 /// and it is 'desirable' to use the type for the given node type. e.g. On x86
2038 /// i16 is legal, but undesirable since i16 instruction encodings are longer
2039 /// and some i16 instructions are slow.
2040 virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const {
2041 // By default, assume all legal types are desirable.
2042 return isTypeLegal(VT);
2045 /// Return true if it is profitable for dag combiner to transform a floating
2046 /// point op of specified opcode to a equivalent op of an integer
2047 /// type. e.g. f32 load -> i32 load can be profitable on ARM.
2048 virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/,
2053 /// This method query the target whether it is beneficial for dag combiner to
2054 /// promote the specified node. If true, it should return the desired
2055 /// promotion type by reference.
2056 virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const {
2060 //===--------------------------------------------------------------------===//
2061 // Lowering methods - These methods must be implemented by targets so that
2062 // the SelectionDAGBuilder code knows how to lower these.
2065 /// This hook must be implemented to lower the incoming (formal) arguments,
2066 /// described by the Ins array, into the specified DAG. The implementation
2067 /// should fill in the InVals array with legal-type argument values, and
2068 /// return the resulting token chain value.
2071 LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/,
2073 const SmallVectorImpl<ISD::InputArg> &/*Ins*/,
2074 SDLoc /*dl*/, SelectionDAG &/*DAG*/,
2075 SmallVectorImpl<SDValue> &/*InVals*/) const {
2076 llvm_unreachable("Not Implemented");
2079 struct ArgListEntry {
2088 bool isInAlloca : 1;
2089 bool isReturned : 1;
2092 ArgListEntry() : isSExt(false), isZExt(false), isInReg(false),
2093 isSRet(false), isNest(false), isByVal(false), isInAlloca(false),
2094 isReturned(false), Alignment(0) { }
2096 void setAttributes(ImmutableCallSite *CS, unsigned AttrIdx);
2098 typedef std::vector<ArgListEntry> ArgListTy;
2100 /// This structure contains all information that is necessary for lowering
2101 /// calls. It is passed to TLI::LowerCallTo when the SelectionDAG builder
2102 /// needs to lower a call, and targets will see this struct in their LowerCall
2104 struct CallLoweringInfo {
2111 bool DoesNotReturn : 1;
2112 bool IsReturnValueUsed : 1;
2114 // IsTailCall should be modified by implementations of
2115 // TargetLowering::LowerCall that perform tail call conversions.
2118 unsigned NumFixedArgs;
2119 CallingConv::ID CallConv;
2124 ImmutableCallSite *CS;
2125 SmallVector<ISD::OutputArg, 32> Outs;
2126 SmallVector<SDValue, 32> OutVals;
2127 SmallVector<ISD::InputArg, 32> Ins;
2130 /// Constructs a call lowering context based on the ImmutableCallSite \p cs.
2131 CallLoweringInfo(SDValue chain, Type *retTy,
2132 FunctionType *FTy, bool isTailCall, SDValue callee,
2133 ArgListTy &args, SelectionDAG &dag, SDLoc dl,
2134 ImmutableCallSite &cs)
2135 : Chain(chain), RetTy(retTy), RetSExt(cs.paramHasAttr(0, Attribute::SExt)),
2136 RetZExt(cs.paramHasAttr(0, Attribute::ZExt)), IsVarArg(FTy->isVarArg()),
2137 IsInReg(cs.paramHasAttr(0, Attribute::InReg)),
2138 DoesNotReturn(cs.doesNotReturn()),
2139 IsReturnValueUsed(!cs.getInstruction()->use_empty()),
2140 IsTailCall(isTailCall), NumFixedArgs(FTy->getNumParams()),
2141 CallConv(cs.getCallingConv()), Callee(callee), Args(args), DAG(dag),
2144 /// Constructs a call lowering context based on the provided call
2146 CallLoweringInfo(SDValue chain, Type *retTy, bool retSExt, bool retZExt,
2147 bool isVarArg, bool isInReg, unsigned numFixedArgs,
2148 CallingConv::ID callConv, bool isTailCall,
2149 bool doesNotReturn, bool isReturnValueUsed, SDValue callee,
2150 ArgListTy &args, SelectionDAG &dag, SDLoc dl)
2151 : Chain(chain), RetTy(retTy), RetSExt(retSExt), RetZExt(retZExt),
2152 IsVarArg(isVarArg), IsInReg(isInReg), DoesNotReturn(doesNotReturn),
2153 IsReturnValueUsed(isReturnValueUsed), IsTailCall(isTailCall),
2154 NumFixedArgs(numFixedArgs), CallConv(callConv), Callee(callee),
2155 Args(args), DAG(dag), DL(dl), CS(nullptr) {}
2158 /// This function lowers an abstract call to a function into an actual call.
2159 /// This returns a pair of operands. The first element is the return value
2160 /// for the function (if RetTy is not VoidTy). The second element is the
2161 /// outgoing token chain. It calls LowerCall to do the actual lowering.
2162 std::pair<SDValue, SDValue> LowerCallTo(CallLoweringInfo &CLI) const;
2164 /// This hook must be implemented to lower calls into the the specified
2165 /// DAG. The outgoing arguments to the call are described by the Outs array,
2166 /// and the values to be returned by the call are described by the Ins
2167 /// array. The implementation should fill in the InVals array with legal-type
2168 /// return values from the call, and return the resulting token chain value.
2170 LowerCall(CallLoweringInfo &/*CLI*/,
2171 SmallVectorImpl<SDValue> &/*InVals*/) const {
2172 llvm_unreachable("Not Implemented");
2175 /// Target-specific cleanup for formal ByVal parameters.
2176 virtual void HandleByVal(CCState *, unsigned &, unsigned) const {}
2178 /// This hook should be implemented to check whether the return values
2179 /// described by the Outs array can fit into the return registers. If false
2180 /// is returned, an sret-demotion is performed.
2181 virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/,
2182 MachineFunction &/*MF*/, bool /*isVarArg*/,
2183 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/,
2184 LLVMContext &/*Context*/) const
2186 // Return true by default to get preexisting behavior.
2190 /// This hook must be implemented to lower outgoing return values, described
2191 /// by the Outs array, into the specified DAG. The implementation should
2192 /// return the resulting token chain value.
2194 LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/,
2196 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/,
2197 const SmallVectorImpl<SDValue> &/*OutVals*/,
2198 SDLoc /*dl*/, SelectionDAG &/*DAG*/) const {
2199 llvm_unreachable("Not Implemented");
2202 /// Return true if result of the specified node is used by a return node
2203 /// only. It also compute and return the input chain for the tail call.
2205 /// This is used to determine whether it is possible to codegen a libcall as
2206 /// tail call at legalization time.
2207 virtual bool isUsedByReturnOnly(SDNode *, SDValue &/*Chain*/) const {
2211 /// Return true if the target may be able emit the call instruction as a tail
2212 /// call. This is used by optimization passes to determine if it's profitable
2213 /// to duplicate return instructions to enable tailcall optimization.
2214 virtual bool mayBeEmittedAsTailCall(CallInst *) const {
2218 /// Return the builtin name for the __builtin___clear_cache intrinsic
2219 /// Default is to invoke the clear cache library call
2220 virtual const char * getClearCacheBuiltinName() const {
2221 return "__clear_cache";
2224 /// Return the register ID of the name passed in. Used by named register
2225 /// global variables extension. There is no target-independent behaviour
2226 /// so the default action is to bail.
2227 virtual unsigned getRegisterByName(const char* RegName, EVT VT) const {
2228 report_fatal_error("Named registers not implemented for this target");
2231 /// Return the type that should be used to zero or sign extend a
2232 /// zeroext/signext integer argument or return value. FIXME: Most C calling
2233 /// convention requires the return type to be promoted, but this is not true
2234 /// all the time, e.g. i1 on x86-64. It is also not necessary for non-C
2235 /// calling conventions. The frontend should handle this and include all of
2236 /// the necessary information.
2237 virtual MVT getTypeForExtArgOrReturn(MVT VT,
2238 ISD::NodeType /*ExtendKind*/) const {
2239 MVT MinVT = getRegisterType(MVT::i32);
2240 return VT.bitsLT(MinVT) ? MinVT : VT;
2243 /// For some targets, an LLVM struct type must be broken down into multiple
2244 /// simple types, but the calling convention specifies that the entire struct
2245 /// must be passed in a block of consecutive registers.
2247 functionArgumentNeedsConsecutiveRegisters(Type *Ty, CallingConv::ID CallConv,
2248 bool isVarArg) const {
2252 /// Returns a 0 terminated array of registers that can be safely used as
2253 /// scratch registers.
2254 virtual const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const {
2258 /// This callback is used to prepare for a volatile or atomic load.
2259 /// It takes a chain node as input and returns the chain for the load itself.
2261 /// Having a callback like this is necessary for targets like SystemZ,
2262 /// which allows a CPU to reuse the result of a previous load indefinitely,
2263 /// even if a cache-coherent store is performed by another CPU. The default
2264 /// implementation does nothing.
2265 virtual SDValue prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL,
2266 SelectionDAG &DAG) const {
2270 /// This callback is invoked by the type legalizer to legalize nodes with an
2271 /// illegal operand type but legal result types. It replaces the
2272 /// LowerOperation callback in the type Legalizer. The reason we can not do
2273 /// away with LowerOperation entirely is that LegalizeDAG isn't yet ready to
2274 /// use this callback.
2276 /// TODO: Consider merging with ReplaceNodeResults.
2278 /// The target places new result values for the node in Results (their number
2279 /// and types must exactly match those of the original return values of
2280 /// the node), or leaves Results empty, which indicates that the node is not
2281 /// to be custom lowered after all.
2282 /// The default implementation calls LowerOperation.
2283 virtual void LowerOperationWrapper(SDNode *N,
2284 SmallVectorImpl<SDValue> &Results,
2285 SelectionDAG &DAG) const;
2287 /// This callback is invoked for operations that are unsupported by the
2288 /// target, which are registered to use 'custom' lowering, and whose defined
2289 /// values are all legal. If the target has no operations that require custom
2290 /// lowering, it need not implement this. The default implementation of this
2292 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
2294 /// This callback is invoked when a node result type is illegal for the
2295 /// target, and the operation was registered to use 'custom' lowering for that
2296 /// result type. The target places new result values for the node in Results
2297 /// (their number and types must exactly match those of the original return
2298 /// values of the node), or leaves Results empty, which indicates that the
2299 /// node is not to be custom lowered after all.
2301 /// If the target has no operations that require custom lowering, it need not
2302 /// implement this. The default implementation aborts.
2303 virtual void ReplaceNodeResults(SDNode * /*N*/,
2304 SmallVectorImpl<SDValue> &/*Results*/,
2305 SelectionDAG &/*DAG*/) const {
2306 llvm_unreachable("ReplaceNodeResults not implemented for this target!");
2309 /// This method returns the name of a target specific DAG node.
2310 virtual const char *getTargetNodeName(unsigned Opcode) const;
2312 /// This method returns a target specific FastISel object, or null if the
2313 /// target does not support "fast" ISel.
2314 virtual FastISel *createFastISel(FunctionLoweringInfo &,
2315 const TargetLibraryInfo *) const {
2320 bool verifyReturnAddressArgumentIsConstant(SDValue Op,
2321 SelectionDAG &DAG) const;
2323 //===--------------------------------------------------------------------===//
2324 // Inline Asm Support hooks
2327 /// This hook allows the target to expand an inline asm call to be explicit
2328 /// llvm code if it wants to. This is useful for turning simple inline asms
2329 /// into LLVM intrinsics, which gives the compiler more information about the
2330 /// behavior of the code.
2331 virtual bool ExpandInlineAsm(CallInst *) const {
2335 enum ConstraintType {
2336 C_Register, // Constraint represents specific register(s).
2337 C_RegisterClass, // Constraint represents any of register(s) in class.
2338 C_Memory, // Memory constraint.
2339 C_Other, // Something else.
2340 C_Unknown // Unsupported constraint.
2343 enum ConstraintWeight {
2345 CW_Invalid = -1, // No match.
2346 CW_Okay = 0, // Acceptable.
2347 CW_Good = 1, // Good weight.
2348 CW_Better = 2, // Better weight.
2349 CW_Best = 3, // Best weight.
2351 // Well-known weights.
2352 CW_SpecificReg = CW_Okay, // Specific register operands.
2353 CW_Register = CW_Good, // Register operands.
2354 CW_Memory = CW_Better, // Memory operands.
2355 CW_Constant = CW_Best, // Constant operand.
2356 CW_Default = CW_Okay // Default or don't know type.
2359 /// This contains information for each constraint that we are lowering.
2360 struct AsmOperandInfo : public InlineAsm::ConstraintInfo {
2361 /// This contains the actual string for the code, like "m". TargetLowering
2362 /// picks the 'best' code from ConstraintInfo::Codes that most closely
2363 /// matches the operand.
2364 std::string ConstraintCode;
2366 /// Information about the constraint code, e.g. Register, RegisterClass,
2367 /// Memory, Other, Unknown.
2368 TargetLowering::ConstraintType ConstraintType;
2370 /// If this is the result output operand or a clobber, this is null,
2371 /// otherwise it is the incoming operand to the CallInst. This gets
2372 /// modified as the asm is processed.
2373 Value *CallOperandVal;
2375 /// The ValueType for the operand value.
2378 /// Return true of this is an input operand that is a matching constraint
2380 bool isMatchingInputConstraint() const;
2382 /// If this is an input matching constraint, this method returns the output
2383 /// operand it matches.
2384 unsigned getMatchedOperand() const;
2386 /// Copy constructor for copying from a ConstraintInfo.
2387 AsmOperandInfo(const InlineAsm::ConstraintInfo &info)
2388 : InlineAsm::ConstraintInfo(info),
2389 ConstraintType(TargetLowering::C_Unknown),
2390 CallOperandVal(nullptr), ConstraintVT(MVT::Other) {
2394 typedef std::vector<AsmOperandInfo> AsmOperandInfoVector;
2396 /// Split up the constraint string from the inline assembly value into the
2397 /// specific constraints and their prefixes, and also tie in the associated
2398 /// operand values. If this returns an empty vector, and if the constraint
2399 /// string itself isn't empty, there was an error parsing.
2400 virtual AsmOperandInfoVector ParseConstraints(ImmutableCallSite CS) const;
2402 /// Examine constraint type and operand type and determine a weight value.
2403 /// The operand object must already have been set up with the operand type.
2404 virtual ConstraintWeight getMultipleConstraintMatchWeight(
2405 AsmOperandInfo &info, int maIndex) const;
2407 /// Examine constraint string and operand type and determine a weight value.
2408 /// The operand object must already have been set up with the operand type.
2409 virtual ConstraintWeight getSingleConstraintMatchWeight(
2410 AsmOperandInfo &info, const char *constraint) const;
2412 /// Determines the constraint code and constraint type to use for the specific
2413 /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType.
2414 /// If the actual operand being passed in is available, it can be passed in as
2415 /// Op, otherwise an empty SDValue can be passed.
2416 virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2418 SelectionDAG *DAG = nullptr) const;
2420 /// Given a constraint, return the type of constraint it is for this target.
2421 virtual ConstraintType getConstraintType(const std::string &Constraint) const;
2423 /// Given a physical register constraint (e.g. {edx}), return the register
2424 /// number and the register class for the register.
2426 /// Given a register class constraint, like 'r', if this corresponds directly
2427 /// to an LLVM register class, return a register of 0 and the register class
2430 /// This should only be used for C_Register constraints. On error, this
2431 /// returns a register number of 0 and a null register class pointer..
2432 virtual std::pair<unsigned, const TargetRegisterClass*>
2433 getRegForInlineAsmConstraint(const std::string &Constraint,
2436 /// Try to replace an X constraint, which matches anything, with another that
2437 /// has more specific requirements based on the type of the corresponding
2438 /// operand. This returns null if there is no replacement to make.
2439 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
2441 /// Lower the specified operand into the Ops vector. If it is invalid, don't
2442 /// add anything to Ops.
2443 virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
2444 std::vector<SDValue> &Ops,
2445 SelectionDAG &DAG) const;
2447 //===--------------------------------------------------------------------===//
2448 // Div utility functions
2450 SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl,
2451 SelectionDAG &DAG) const;
2452 SDValue BuildSDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
2453 bool IsAfterLegalization,
2454 std::vector<SDNode *> *Created) const;
2455 SDValue BuildUDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
2456 bool IsAfterLegalization,
2457 std::vector<SDNode *> *Created) const;
2459 //===--------------------------------------------------------------------===//
2460 // Legalization utility functions
2463 /// Expand a MUL into two nodes. One that computes the high bits of
2464 /// the result and one that computes the low bits.
2465 /// \param HiLoVT The value type to use for the Lo and Hi nodes.
2466 /// \param LL Low bits of the LHS of the MUL. You can use this parameter
2467 /// if you want to control how low bits are extracted from the LHS.
2468 /// \param LH High bits of the LHS of the MUL. See LL for meaning.
2469 /// \param RL Low bits of the RHS of the MUL. See LL for meaning
2470 /// \param RH High bits of the RHS of the MUL. See LL for meaning.
2471 /// \returns true if the node has been expanded. false if it has not
2472 bool expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT,
2473 SelectionDAG &DAG, SDValue LL = SDValue(),
2474 SDValue LH = SDValue(), SDValue RL = SDValue(),
2475 SDValue RH = SDValue()) const;
2477 //===--------------------------------------------------------------------===//
2478 // Instruction Emitting Hooks
2481 /// This method should be implemented by targets that mark instructions with
2482 /// the 'usesCustomInserter' flag. These instructions are special in various
2483 /// ways, which require special support to insert. The specified MachineInstr
2484 /// is created but not inserted into any basic blocks, and this method is
2485 /// called to expand it into a sequence of instructions, potentially also
2486 /// creating new basic blocks and control flow.
2487 virtual MachineBasicBlock *
2488 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const;
2490 /// This method should be implemented by targets that mark instructions with
2491 /// the 'hasPostISelHook' flag. These instructions must be adjusted after
2492 /// instruction selection by target hooks. e.g. To fill in optional defs for
2493 /// ARM 's' setting instructions.
2495 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const;
2498 /// Given an LLVM IR type and return type attributes, compute the return value
2499 /// EVTs and flags, and optionally also the offsets, if the return value is
2500 /// being lowered to memory.
2501 void GetReturnInfo(Type* ReturnType, AttributeSet attr,
2502 SmallVectorImpl<ISD::OutputArg> &Outs,
2503 const TargetLowering &TLI);
2505 } // end llvm namespace