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/MC/MCRegisterInfo.h"
36 #include "llvm/Target/TargetCallingConv.h"
37 #include "llvm/Target/TargetMachine.h"
46 class FunctionLoweringInfo;
47 class ImmutableCallSite;
49 class MachineBasicBlock;
50 class MachineFunction;
52 class MachineJumpTableInfo;
57 template<typename T> class SmallVectorImpl;
59 class TargetRegisterClass;
60 class TargetLibraryInfo;
61 class TargetLoweringObjectFile;
66 None, // No preference
67 Source, // Follow source order.
68 RegPressure, // Scheduling for lowest register pressure.
69 Hybrid, // Scheduling for both latency and register pressure.
70 ILP, // Scheduling for ILP in low register pressure mode.
71 VLIW // Scheduling for VLIW targets.
75 /// This base class for TargetLowering contains the SelectionDAG-independent
76 /// parts that can be used from the rest of CodeGen.
77 class TargetLoweringBase {
78 TargetLoweringBase(const TargetLoweringBase&) LLVM_DELETED_FUNCTION;
79 void operator=(const TargetLoweringBase&) LLVM_DELETED_FUNCTION;
82 /// This enum indicates whether operations are valid for a target, and if not,
83 /// what action should be used to make them valid.
85 Legal, // The target natively supports this operation.
86 Promote, // This operation should be executed in a larger type.
87 Expand, // Try to expand this to other ops, otherwise use a libcall.
88 Custom // Use the LowerOperation hook to implement custom lowering.
91 /// This enum indicates whether a types are legal for a target, and if not,
92 /// what action should be used to make them valid.
93 enum LegalizeTypeAction {
94 TypeLegal, // The target natively supports this type.
95 TypePromoteInteger, // Replace this integer with a larger one.
96 TypeExpandInteger, // Split this integer into two of half the size.
97 TypeSoftenFloat, // Convert this float to a same size integer type.
98 TypeExpandFloat, // Split this float into two of half the size.
99 TypeScalarizeVector, // Replace this one-element vector with its element.
100 TypeSplitVector, // Split this vector into two of half the size.
101 TypeWidenVector // This vector should be widened into a larger vector.
104 /// LegalizeKind holds the legalization kind that needs to happen to EVT
105 /// in order to type-legalize it.
106 typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind;
108 /// Enum that describes how the target represents true/false values.
109 enum BooleanContent {
110 UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage.
111 ZeroOrOneBooleanContent, // All bits zero except for bit 0.
112 ZeroOrNegativeOneBooleanContent // All bits equal to bit 0.
115 /// Enum that describes what type of support for selects the target has.
116 enum SelectSupportKind {
117 ScalarValSelect, // The target supports scalar selects (ex: cmov).
118 ScalarCondVectorVal, // The target supports selects with a scalar condition
119 // and vector values (ex: cmov).
120 VectorMaskSelect // The target supports vector selects with a vector
121 // mask (ex: x86 blends).
124 static ISD::NodeType getExtendForContent(BooleanContent Content) {
126 case UndefinedBooleanContent:
127 // Extend by adding rubbish bits.
128 return ISD::ANY_EXTEND;
129 case ZeroOrOneBooleanContent:
130 // Extend by adding zero bits.
131 return ISD::ZERO_EXTEND;
132 case ZeroOrNegativeOneBooleanContent:
133 // Extend by copying the sign bit.
134 return ISD::SIGN_EXTEND;
136 llvm_unreachable("Invalid content kind");
139 /// NOTE: The constructor takes ownership of TLOF.
140 explicit TargetLoweringBase(const TargetMachine &TM,
141 const TargetLoweringObjectFile *TLOF);
142 virtual ~TargetLoweringBase();
145 /// \brief Initialize all of the actions to default values.
149 const TargetMachine &getTargetMachine() const { return TM; }
150 const DataLayout *getDataLayout() const { return DL; }
151 const TargetLoweringObjectFile &getObjFileLowering() const { return TLOF; }
153 bool isBigEndian() const { return !IsLittleEndian; }
154 bool isLittleEndian() const { return IsLittleEndian; }
156 /// Return the pointer type for the given address space, defaults to
157 /// the pointer type from the data layout.
158 /// FIXME: The default needs to be removed once all the code is updated.
159 virtual MVT getPointerTy(uint32_t /*AS*/ = 0) const;
160 unsigned getPointerSizeInBits(uint32_t AS = 0) const;
161 unsigned getPointerTypeSizeInBits(Type *Ty) const;
162 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const;
164 EVT getShiftAmountTy(EVT LHSTy) const;
166 /// Returns the type to be used for the index operand of:
167 /// ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT,
168 /// ISD::INSERT_SUBVECTOR, and ISD::EXTRACT_SUBVECTOR
169 virtual MVT getVectorIdxTy() const {
170 return getPointerTy();
173 /// Return true if the select operation is expensive for this target.
174 bool isSelectExpensive() const { return SelectIsExpensive; }
176 virtual bool isSelectSupported(SelectSupportKind /*kind*/) const {
180 /// Return true if multiple condition registers are available.
181 bool hasMultipleConditionRegisters() const {
182 return HasMultipleConditionRegisters;
185 /// Return true if the target has BitExtract instructions.
186 bool hasExtractBitsInsn() const { return HasExtractBitsInsn; }
188 /// Return the preferred vector type legalization action.
189 virtual TargetLoweringBase::LegalizeTypeAction
190 getPreferredVectorAction(EVT VT) const {
191 // The default action for one element vectors is to scalarize
192 if (VT.getVectorNumElements() == 1)
193 return TypeScalarizeVector;
194 // The default action for other vectors is to promote
195 return TypePromoteInteger;
198 // There are two general methods for expanding a BUILD_VECTOR node:
199 // 1. Use SCALAR_TO_VECTOR on the defined scalar values and then shuffle
201 // 2. Build the vector on the stack and then load it.
202 // If this function returns true, then method (1) will be used, subject to
203 // the constraint that all of the necessary shuffles are legal (as determined
204 // by isShuffleMaskLegal). If this function returns false, then method (2) is
205 // always used. The vector type, and the number of defined values, are
208 shouldExpandBuildVectorWithShuffles(EVT /* VT */,
209 unsigned DefinedValues) const {
210 return DefinedValues < 3;
213 /// Return true if integer divide is usually cheaper than a sequence of
214 /// several shifts, adds, and multiplies for this target.
215 bool isIntDivCheap() const { return IntDivIsCheap; }
217 /// Returns true if target has indicated at least one type should be bypassed.
218 bool isSlowDivBypassed() const { return !BypassSlowDivWidths.empty(); }
220 /// Returns map of slow types for division or remainder with corresponding
222 const DenseMap<unsigned int, unsigned int> &getBypassSlowDivWidths() const {
223 return BypassSlowDivWidths;
226 /// Return true if pow2 div is cheaper than a chain of srl/add/sra.
227 bool isPow2DivCheap() const { return Pow2DivIsCheap; }
229 /// Return true if Flow Control is an expensive operation that should be
231 bool isJumpExpensive() const { return JumpIsExpensive; }
233 /// Return true if selects are only cheaper than branches if the branch is
234 /// unlikely to be predicted right.
235 bool isPredictableSelectExpensive() const {
236 return PredictableSelectIsExpensive;
239 /// isLoadBitCastBeneficial() - Return true if the following transform
241 /// fold (conv (load x)) -> (load (conv*)x)
242 /// On architectures that don't natively support some vector loads efficiently,
243 /// casting the load to a smaller vector of larger types and loading
244 /// is more efficient, however, this can be undone by optimizations in
246 virtual bool isLoadBitCastBeneficial(EVT /* Load */, EVT /* Bitcast */) const {
250 /// \brief Return if the target supports combining a
253 /// %andResult = and %val1, #imm-with-one-bit-set;
254 /// %icmpResult = icmp %andResult, 0
255 /// br i1 %icmpResult, label %dest1, label %dest2
257 /// into a single machine instruction of a form like:
259 /// brOnBitSet %register, #bitNumber, dest
261 bool isMaskAndBranchFoldingLegal() const {
262 return MaskAndBranchFoldingIsLegal;
265 /// Return the ValueType of the result of SETCC operations. Also used to
266 /// obtain the target's preferred type for the condition operand of SELECT and
267 /// BRCOND nodes. In the case of BRCOND the argument passed is MVT::Other
268 /// since there are no other operands to get a type hint from.
269 virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
271 /// Return the ValueType for comparison libcalls. Comparions libcalls include
272 /// floating point comparion calls, and Ordered/Unordered check calls on
273 /// floating point numbers.
275 MVT::SimpleValueType getCmpLibcallReturnType() const;
277 /// For targets without i1 registers, this gives the nature of the high-bits
278 /// of boolean values held in types wider than i1.
280 /// "Boolean values" are special true/false values produced by nodes like
281 /// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND.
282 /// Not to be confused with general values promoted from i1. Some cpus
283 /// distinguish between vectors of boolean and scalars; the isVec parameter
284 /// selects between the two kinds. For example on X86 a scalar boolean should
285 /// be zero extended from i1, while the elements of a vector of booleans
286 /// should be sign extended from i1.
288 /// Some cpus also treat floating point types the same way as they treat
289 /// vectors instead of the way they treat scalars.
290 BooleanContent getBooleanContents(bool isVec, bool isFloat) const {
292 return BooleanVectorContents;
293 return isFloat ? BooleanFloatContents : BooleanContents;
296 BooleanContent getBooleanContents(EVT Type) const {
297 return getBooleanContents(Type.isVector(), Type.isFloatingPoint());
300 /// Return target scheduling preference.
301 Sched::Preference getSchedulingPreference() const {
302 return SchedPreferenceInfo;
305 /// Some scheduler, e.g. hybrid, can switch to different scheduling heuristics
306 /// for different nodes. This function returns the preference (or none) for
308 virtual Sched::Preference getSchedulingPreference(SDNode *) const {
312 /// Return the register class that should be used for the specified value
314 virtual const TargetRegisterClass *getRegClassFor(MVT VT) const {
315 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
316 assert(RC && "This value type is not natively supported!");
320 /// Return the 'representative' register class for the specified value
323 /// The 'representative' register class is the largest legal super-reg
324 /// register class for the register class of the value type. For example, on
325 /// i386 the rep register class for i8, i16, and i32 are GR32; while the rep
326 /// register class is GR64 on x86_64.
327 virtual const TargetRegisterClass *getRepRegClassFor(MVT VT) const {
328 const TargetRegisterClass *RC = RepRegClassForVT[VT.SimpleTy];
332 /// Return the cost of the 'representative' register class for the specified
334 virtual uint8_t getRepRegClassCostFor(MVT VT) const {
335 return RepRegClassCostForVT[VT.SimpleTy];
338 /// Return true if the target has native support for the specified value type.
339 /// This means that it has a register that directly holds it without
340 /// promotions or expansions.
341 bool isTypeLegal(EVT VT) const {
342 assert(!VT.isSimple() ||
343 (unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT));
344 return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != nullptr;
347 class ValueTypeActionImpl {
348 /// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum
349 /// that indicates how instruction selection should deal with the type.
350 uint8_t ValueTypeActions[MVT::LAST_VALUETYPE];
353 ValueTypeActionImpl() {
354 std::fill(std::begin(ValueTypeActions), std::end(ValueTypeActions), 0);
357 LegalizeTypeAction getTypeAction(MVT VT) const {
358 return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy];
361 void setTypeAction(MVT VT, LegalizeTypeAction Action) {
362 unsigned I = VT.SimpleTy;
363 ValueTypeActions[I] = Action;
367 const ValueTypeActionImpl &getValueTypeActions() const {
368 return ValueTypeActions;
371 /// Return how we should legalize values of this type, either it is already
372 /// legal (return 'Legal') or we need to promote it to a larger type (return
373 /// 'Promote'), or we need to expand it into multiple registers of smaller
374 /// integer type (return 'Expand'). 'Custom' is not an option.
375 LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const {
376 return getTypeConversion(Context, VT).first;
378 LegalizeTypeAction getTypeAction(MVT VT) const {
379 return ValueTypeActions.getTypeAction(VT);
382 /// For types supported by the target, this is an identity function. For
383 /// types that must be promoted to larger types, this returns the larger type
384 /// to promote to. For integer types that are larger than the largest integer
385 /// register, this contains one step in the expansion to get to the smaller
386 /// register. For illegal floating point types, this returns the integer type
388 EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const {
389 return getTypeConversion(Context, VT).second;
392 /// For types supported by the target, this is an identity function. For
393 /// types that must be expanded (i.e. integer types that are larger than the
394 /// largest integer register or illegal floating point types), this returns
395 /// the largest legal type it will be expanded to.
396 EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const {
397 assert(!VT.isVector());
399 switch (getTypeAction(Context, VT)) {
402 case TypeExpandInteger:
403 VT = getTypeToTransformTo(Context, VT);
406 llvm_unreachable("Type is not legal nor is it to be expanded!");
411 /// Vector types are broken down into some number of legal first class types.
412 /// For example, EVT::v8f32 maps to 2 EVT::v4f32 with Altivec or SSE1, or 8
413 /// promoted EVT::f64 values with the X86 FP stack. Similarly, EVT::v2i64
414 /// turns into 4 EVT::i32 values with both PPC and X86.
416 /// This method returns the number of registers needed, and the VT for each
417 /// register. It also returns the VT and quantity of the intermediate values
418 /// before they are promoted/expanded.
419 unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
421 unsigned &NumIntermediates,
422 MVT &RegisterVT) const;
424 struct IntrinsicInfo {
425 unsigned opc; // target opcode
426 EVT memVT; // memory VT
427 const Value* ptrVal; // value representing memory location
428 int offset; // offset off of ptrVal
429 unsigned align; // alignment
430 bool vol; // is volatile?
431 bool readMem; // reads memory?
432 bool writeMem; // writes memory?
435 /// Given an intrinsic, checks if on the target the intrinsic will need to map
436 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns
437 /// true and store the intrinsic information into the IntrinsicInfo that was
438 /// passed to the function.
439 virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &,
440 unsigned /*Intrinsic*/) const {
444 /// Returns true if the target can instruction select the specified FP
445 /// immediate natively. If false, the legalizer will materialize the FP
446 /// immediate as a load from a constant pool.
447 virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const {
451 /// Targets can use this to indicate that they only support *some*
452 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a
453 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to be
455 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/,
460 /// Returns true if the operation can trap for the value type.
462 /// VT must be a legal type. By default, we optimistically assume most
463 /// operations don't trap except for divide and remainder.
464 virtual bool canOpTrap(unsigned Op, EVT VT) const;
466 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to
467 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to replace
468 /// a VAND with a constant pool entry.
469 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/,
474 /// Return how this operation should be treated: either it is legal, needs to
475 /// be promoted to a larger size, needs to be expanded to some other code
476 /// sequence, or the target has a custom expander for it.
477 LegalizeAction getOperationAction(unsigned Op, EVT VT) const {
478 if (VT.isExtended()) return Expand;
479 // If a target-specific SDNode requires legalization, require the target
480 // to provide custom legalization for it.
481 if (Op > array_lengthof(OpActions[0])) return Custom;
482 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy;
483 return (LegalizeAction)OpActions[I][Op];
486 /// Return true if the specified operation is legal on this target or can be
487 /// made legal with custom lowering. This is used to help guide high-level
488 /// lowering decisions.
489 bool isOperationLegalOrCustom(unsigned Op, EVT VT) const {
490 return (VT == MVT::Other || isTypeLegal(VT)) &&
491 (getOperationAction(Op, VT) == Legal ||
492 getOperationAction(Op, VT) == Custom);
495 /// Return true if the specified operation is legal on this target or can be
496 /// made legal using promotion. This is used to help guide high-level lowering
498 bool isOperationLegalOrPromote(unsigned Op, EVT VT) const {
499 return (VT == MVT::Other || isTypeLegal(VT)) &&
500 (getOperationAction(Op, VT) == Legal ||
501 getOperationAction(Op, VT) == Promote);
504 /// Return true if the specified operation is illegal on this target or
505 /// unlikely to be made legal with custom lowering. This is used to help guide
506 /// high-level lowering decisions.
507 bool isOperationExpand(unsigned Op, EVT VT) const {
508 return (!isTypeLegal(VT) || getOperationAction(Op, VT) == Expand);
511 /// Return true if the specified operation is legal on this target.
512 bool isOperationLegal(unsigned Op, EVT VT) const {
513 return (VT == MVT::Other || isTypeLegal(VT)) &&
514 getOperationAction(Op, VT) == Legal;
517 /// Return how this load with extension should be treated: either it is legal,
518 /// needs to be promoted to a larger size, needs to be expanded to some other
519 /// code sequence, or the target has a custom expander for it.
520 LegalizeAction getLoadExtAction(unsigned ExtType, EVT VT) const {
521 if (VT.isExtended()) return Expand;
522 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy;
523 assert(ExtType < ISD::LAST_LOADEXT_TYPE && I < MVT::LAST_VALUETYPE &&
524 "Table isn't big enough!");
525 return (LegalizeAction)LoadExtActions[I][ExtType];
528 /// Return true if the specified load with extension is legal on this target.
529 bool isLoadExtLegal(unsigned ExtType, EVT VT) const {
530 return VT.isSimple() &&
531 getLoadExtAction(ExtType, VT.getSimpleVT()) == Legal;
534 /// Return how this store with truncation should be treated: either it is
535 /// legal, needs to be promoted to a larger size, needs to be expanded to some
536 /// other code sequence, or the target has a custom expander for it.
537 LegalizeAction getTruncStoreAction(EVT ValVT, EVT MemVT) const {
538 if (ValVT.isExtended() || MemVT.isExtended()) return Expand;
539 unsigned ValI = (unsigned) ValVT.getSimpleVT().SimpleTy;
540 unsigned MemI = (unsigned) MemVT.getSimpleVT().SimpleTy;
541 assert(ValI < MVT::LAST_VALUETYPE && MemI < MVT::LAST_VALUETYPE &&
542 "Table isn't big enough!");
543 return (LegalizeAction)TruncStoreActions[ValI][MemI];
546 /// Return true if the specified store with truncation is legal on this
548 bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const {
549 return isTypeLegal(ValVT) && MemVT.isSimple() &&
550 getTruncStoreAction(ValVT.getSimpleVT(), MemVT.getSimpleVT()) == Legal;
553 /// Return how the indexed load should be treated: either it is legal, needs
554 /// to be promoted to a larger size, needs to be expanded to some other code
555 /// sequence, or the target has a custom expander for it.
557 getIndexedLoadAction(unsigned IdxMode, MVT VT) const {
558 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE &&
559 "Table isn't big enough!");
560 unsigned Ty = (unsigned)VT.SimpleTy;
561 return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4);
564 /// Return true if the specified indexed load is legal on this target.
565 bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const {
566 return VT.isSimple() &&
567 (getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Legal ||
568 getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Custom);
571 /// Return how the indexed store should be treated: either it is legal, needs
572 /// to be promoted to a larger size, needs to be expanded to some other code
573 /// sequence, or the target has a custom expander for it.
575 getIndexedStoreAction(unsigned IdxMode, MVT VT) const {
576 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE &&
577 "Table isn't big enough!");
578 unsigned Ty = (unsigned)VT.SimpleTy;
579 return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f);
582 /// Return true if the specified indexed load is legal on this target.
583 bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const {
584 return VT.isSimple() &&
585 (getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Legal ||
586 getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Custom);
589 /// Return how the condition code should be treated: either it is legal, needs
590 /// to be expanded to some other code sequence, or the target has a custom
593 getCondCodeAction(ISD::CondCode CC, MVT VT) const {
594 assert((unsigned)CC < array_lengthof(CondCodeActions) &&
595 ((unsigned)VT.SimpleTy >> 4) < array_lengthof(CondCodeActions[0]) &&
596 "Table isn't big enough!");
597 // See setCondCodeAction for how this is encoded.
598 uint32_t Shift = 2 * (VT.SimpleTy & 0xF);
599 uint32_t Value = CondCodeActions[CC][VT.SimpleTy >> 4];
600 LegalizeAction Action = (LegalizeAction) ((Value >> Shift) & 0x3);
601 assert(Action != Promote && "Can't promote condition code!");
605 /// Return true if the specified condition code is legal on this target.
606 bool isCondCodeLegal(ISD::CondCode CC, MVT VT) const {
608 getCondCodeAction(CC, VT) == Legal ||
609 getCondCodeAction(CC, VT) == Custom;
613 /// If the action for this operation is to promote, this method returns the
614 /// ValueType to promote to.
615 MVT getTypeToPromoteTo(unsigned Op, MVT VT) const {
616 assert(getOperationAction(Op, VT) == Promote &&
617 "This operation isn't promoted!");
619 // See if this has an explicit type specified.
620 std::map<std::pair<unsigned, MVT::SimpleValueType>,
621 MVT::SimpleValueType>::const_iterator PTTI =
622 PromoteToType.find(std::make_pair(Op, VT.SimpleTy));
623 if (PTTI != PromoteToType.end()) return PTTI->second;
625 assert((VT.isInteger() || VT.isFloatingPoint()) &&
626 "Cannot autopromote this type, add it with AddPromotedToType.");
630 NVT = (MVT::SimpleValueType)(NVT.SimpleTy+1);
631 assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid &&
632 "Didn't find type to promote to!");
633 } while (!isTypeLegal(NVT) ||
634 getOperationAction(Op, NVT) == Promote);
638 /// Return the EVT corresponding to this LLVM type. This is fixed by the LLVM
639 /// operations except for the pointer size. If AllowUnknown is true, this
640 /// will return MVT::Other for types with no EVT counterpart (e.g. structs),
641 /// otherwise it will assert.
642 EVT getValueType(Type *Ty, bool AllowUnknown = false) const {
643 // Lower scalar pointers to native pointer types.
644 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
645 return getPointerTy(PTy->getAddressSpace());
647 if (Ty->isVectorTy()) {
648 VectorType *VTy = cast<VectorType>(Ty);
649 Type *Elm = VTy->getElementType();
650 // Lower vectors of pointers to native pointer types.
651 if (PointerType *PT = dyn_cast<PointerType>(Elm)) {
652 EVT PointerTy(getPointerTy(PT->getAddressSpace()));
653 Elm = PointerTy.getTypeForEVT(Ty->getContext());
656 return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false),
657 VTy->getNumElements());
659 return EVT::getEVT(Ty, AllowUnknown);
662 /// Return the MVT corresponding to this LLVM type. See getValueType.
663 MVT getSimpleValueType(Type *Ty, bool AllowUnknown = false) const {
664 return getValueType(Ty, AllowUnknown).getSimpleVT();
667 /// Return the desired alignment for ByVal or InAlloca aggregate function
668 /// arguments in the caller parameter area. This is the actual alignment, not
670 virtual unsigned getByValTypeAlignment(Type *Ty) const;
672 /// Return the type of registers that this ValueType will eventually require.
673 MVT getRegisterType(MVT VT) const {
674 assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT));
675 return RegisterTypeForVT[VT.SimpleTy];
678 /// Return the type of registers that this ValueType will eventually require.
679 MVT getRegisterType(LLVMContext &Context, EVT VT) const {
681 assert((unsigned)VT.getSimpleVT().SimpleTy <
682 array_lengthof(RegisterTypeForVT));
683 return RegisterTypeForVT[VT.getSimpleVT().SimpleTy];
688 unsigned NumIntermediates;
689 (void)getVectorTypeBreakdown(Context, VT, VT1,
690 NumIntermediates, RegisterVT);
693 if (VT.isInteger()) {
694 return getRegisterType(Context, getTypeToTransformTo(Context, VT));
696 llvm_unreachable("Unsupported extended type!");
699 /// Return the number of registers that this ValueType will eventually
702 /// This is one for any types promoted to live in larger registers, but may be
703 /// more than one for types (like i64) that are split into pieces. For types
704 /// like i140, which are first promoted then expanded, it is the number of
705 /// registers needed to hold all the bits of the original type. For an i140
706 /// on a 32 bit machine this means 5 registers.
707 unsigned getNumRegisters(LLVMContext &Context, EVT VT) const {
709 assert((unsigned)VT.getSimpleVT().SimpleTy <
710 array_lengthof(NumRegistersForVT));
711 return NumRegistersForVT[VT.getSimpleVT().SimpleTy];
716 unsigned NumIntermediates;
717 return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2);
719 if (VT.isInteger()) {
720 unsigned BitWidth = VT.getSizeInBits();
721 unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits();
722 return (BitWidth + RegWidth - 1) / RegWidth;
724 llvm_unreachable("Unsupported extended type!");
727 /// If true, then instruction selection should seek to shrink the FP constant
728 /// of the specified type to a smaller type in order to save space and / or
730 virtual bool ShouldShrinkFPConstant(EVT) const { return true; }
732 /// When splitting a value of the specified type into parts, does the Lo
733 /// or Hi part come first? This usually follows the endianness, except
734 /// for ppcf128, where the Hi part always comes first.
735 bool hasBigEndianPartOrdering(EVT VT) const {
736 return isBigEndian() || VT == MVT::ppcf128;
739 /// If true, the target has custom DAG combine transformations that it can
740 /// perform for the specified node.
741 bool hasTargetDAGCombine(ISD::NodeType NT) const {
742 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
743 return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7));
746 /// \brief Get maximum # of store operations permitted for llvm.memset
748 /// This function returns the maximum number of store operations permitted
749 /// to replace a call to llvm.memset. The value is set by the target at the
750 /// performance threshold for such a replacement. If OptSize is true,
751 /// return the limit for functions that have OptSize attribute.
752 unsigned getMaxStoresPerMemset(bool OptSize) const {
753 return OptSize ? MaxStoresPerMemsetOptSize : MaxStoresPerMemset;
756 /// \brief Get maximum # of store operations permitted for llvm.memcpy
758 /// This function returns the maximum number of store operations permitted
759 /// to replace a call to llvm.memcpy. The value is set by the target at the
760 /// performance threshold for such a replacement. If OptSize is true,
761 /// return the limit for functions that have OptSize attribute.
762 unsigned getMaxStoresPerMemcpy(bool OptSize) const {
763 return OptSize ? MaxStoresPerMemcpyOptSize : MaxStoresPerMemcpy;
766 /// \brief Get maximum # of store operations permitted for llvm.memmove
768 /// This function returns the maximum number of store operations permitted
769 /// to replace a call to llvm.memmove. The value is set by the target at the
770 /// performance threshold for such a replacement. If OptSize is true,
771 /// return the limit for functions that have OptSize attribute.
772 unsigned getMaxStoresPerMemmove(bool OptSize) const {
773 return OptSize ? MaxStoresPerMemmoveOptSize : MaxStoresPerMemmove;
776 /// \brief Determine if the target supports unaligned memory accesses.
778 /// This function returns true if the target allows unaligned memory accesses
779 /// of the specified type in the given address space. If true, it also returns
780 /// whether the unaligned memory access is "fast" in the third argument by
781 /// reference. This is used, for example, in situations where an array
782 /// copy/move/set is converted to a sequence of store operations. Its use
783 /// helps to ensure that such replacements don't generate code that causes an
784 /// alignment error (trap) on the target machine.
785 virtual bool allowsUnalignedMemoryAccesses(EVT,
786 unsigned AddrSpace = 0,
787 bool * /*Fast*/ = nullptr) const {
791 /// Returns the target specific optimal type for load and store operations as
792 /// a result of memset, memcpy, and memmove lowering.
794 /// If DstAlign is zero that means it's safe to destination alignment can
795 /// satisfy any constraint. Similarly if SrcAlign is zero it means there isn't
796 /// a need to check it against alignment requirement, probably because the
797 /// source does not need to be loaded. If 'IsMemset' is true, that means it's
798 /// expanding a memset. If 'ZeroMemset' is true, that means it's a memset of
799 /// zero. 'MemcpyStrSrc' indicates whether the memcpy source is constant so it
800 /// does not need to be loaded. It returns EVT::Other if the type should be
801 /// determined using generic target-independent logic.
802 virtual EVT getOptimalMemOpType(uint64_t /*Size*/,
803 unsigned /*DstAlign*/, unsigned /*SrcAlign*/,
806 bool /*MemcpyStrSrc*/,
807 MachineFunction &/*MF*/) const {
811 /// Returns true if it's safe to use load / store of the specified type to
812 /// expand memcpy / memset inline.
814 /// This is mostly true for all types except for some special cases. For
815 /// example, on X86 targets without SSE2 f64 load / store are done with fldl /
816 /// fstpl which also does type conversion. Note the specified type doesn't
817 /// have to be legal as the hook is used before type legalization.
818 virtual bool isSafeMemOpType(MVT /*VT*/) const { return true; }
820 /// Determine if we should use _setjmp or setjmp to implement llvm.setjmp.
821 bool usesUnderscoreSetJmp() const {
822 return UseUnderscoreSetJmp;
825 /// Determine if we should use _longjmp or longjmp to implement llvm.longjmp.
826 bool usesUnderscoreLongJmp() const {
827 return UseUnderscoreLongJmp;
830 /// Return whether the target can generate code for jump tables.
831 bool supportJumpTables() const {
832 return SupportJumpTables;
835 /// Return integer threshold on number of blocks to use jump tables rather
836 /// than if sequence.
837 int getMinimumJumpTableEntries() const {
838 return MinimumJumpTableEntries;
841 /// If a physical register, this specifies the register that
842 /// llvm.savestack/llvm.restorestack should save and restore.
843 unsigned getStackPointerRegisterToSaveRestore() const {
844 return StackPointerRegisterToSaveRestore;
847 /// If a physical register, this returns the register that receives the
848 /// exception address on entry to a landing pad.
849 unsigned getExceptionPointerRegister() const {
850 return ExceptionPointerRegister;
853 /// If a physical register, this returns the register that receives the
854 /// exception typeid on entry to a landing pad.
855 unsigned getExceptionSelectorRegister() const {
856 return ExceptionSelectorRegister;
859 /// Returns the target's jmp_buf size in bytes (if never set, the default is
861 unsigned getJumpBufSize() const {
865 /// Returns the target's jmp_buf alignment in bytes (if never set, the default
867 unsigned getJumpBufAlignment() const {
868 return JumpBufAlignment;
871 /// Return the minimum stack alignment of an argument.
872 unsigned getMinStackArgumentAlignment() const {
873 return MinStackArgumentAlignment;
876 /// Return the minimum function alignment.
877 unsigned getMinFunctionAlignment() const {
878 return MinFunctionAlignment;
881 /// Return the preferred function alignment.
882 unsigned getPrefFunctionAlignment() const {
883 return PrefFunctionAlignment;
886 /// Return the preferred loop alignment.
887 unsigned getPrefLoopAlignment() const {
888 return PrefLoopAlignment;
891 /// Return whether the DAG builder should automatically insert fences and
892 /// reduce ordering for atomics.
893 bool getInsertFencesForAtomic() const {
894 return InsertFencesForAtomic;
897 /// Return true if the target stores stack protector cookies at a fixed offset
898 /// in some non-standard address space, and populates the address space and
899 /// offset as appropriate.
900 virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/,
901 unsigned &/*Offset*/) const {
905 /// Returns the maximal possible offset which can be used for loads / stores
907 virtual unsigned getMaximalGlobalOffset() const {
911 /// Returns true if a cast between SrcAS and DestAS is a noop.
912 virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const {
916 //===--------------------------------------------------------------------===//
917 /// \name Helpers for TargetTransformInfo implementations
920 /// Get the ISD node that corresponds to the Instruction class opcode.
921 int InstructionOpcodeToISD(unsigned Opcode) const;
923 /// Estimate the cost of type-legalization and the legalized type.
924 std::pair<unsigned, MVT> getTypeLegalizationCost(Type *Ty) const;
928 //===--------------------------------------------------------------------===//
929 /// \name Helpers for load-linked/store-conditional atomic expansion.
932 /// Perform a load-linked operation on Addr, returning a "Value *" with the
933 /// corresponding pointee type. This may entail some non-trivial operations to
934 /// truncate or reconstruct types that will be illegal in the backend. See
935 /// ARMISelLowering for an example implementation.
936 virtual Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
937 AtomicOrdering Ord) const {
938 llvm_unreachable("Load linked unimplemented on this target");
941 /// Perform a store-conditional operation to Addr. Return the status of the
942 /// store. This should be 0 if the store succeeded, non-zero otherwise.
943 virtual Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val,
944 Value *Addr, AtomicOrdering Ord) const {
945 llvm_unreachable("Store conditional unimplemented on this target");
948 /// Return true if the given (atomic) instruction should be expanded by the
949 /// IR-level AtomicExpandLoadLinked pass into a loop involving
950 /// load-linked/store-conditional pairs. Atomic stores will be expanded in the
951 /// same way as "atomic xchg" operations which ignore their output if needed.
952 virtual bool shouldExpandAtomicInIR(Instruction *Inst) const {
957 //===--------------------------------------------------------------------===//
958 // TargetLowering Configuration Methods - These methods should be invoked by
959 // the derived class constructor to configure this object for the target.
962 /// \brief Reset the operation actions based on target options.
963 virtual void resetOperationActions() {}
966 /// Specify how the target extends the result of integer and floating point
967 /// boolean values from i1 to a wider type. See getBooleanContents.
968 void setBooleanContents(BooleanContent Ty) {
969 BooleanContents = Ty;
970 BooleanFloatContents = Ty;
973 /// Specify how the target extends the result of integer and floating point
974 /// boolean values from i1 to a wider type. See getBooleanContents.
975 void setBooleanContents(BooleanContent IntTy, BooleanContent FloatTy) {
976 BooleanContents = IntTy;
977 BooleanFloatContents = FloatTy;
980 /// Specify how the target extends the result of a vector boolean value from a
981 /// vector of i1 to a wider type. See getBooleanContents.
982 void setBooleanVectorContents(BooleanContent Ty) {
983 BooleanVectorContents = Ty;
986 /// Specify the target scheduling preference.
987 void setSchedulingPreference(Sched::Preference Pref) {
988 SchedPreferenceInfo = Pref;
991 /// Indicate whether this target prefers to use _setjmp to implement
992 /// llvm.setjmp or the version without _. Defaults to false.
993 void setUseUnderscoreSetJmp(bool Val) {
994 UseUnderscoreSetJmp = Val;
997 /// Indicate whether this target prefers to use _longjmp to implement
998 /// llvm.longjmp or the version without _. Defaults to false.
999 void setUseUnderscoreLongJmp(bool Val) {
1000 UseUnderscoreLongJmp = Val;
1003 /// Indicate whether the target can generate code for jump tables.
1004 void setSupportJumpTables(bool Val) {
1005 SupportJumpTables = Val;
1008 /// Indicate the number of blocks to generate jump tables rather than if
1010 void setMinimumJumpTableEntries(int Val) {
1011 MinimumJumpTableEntries = Val;
1014 /// If set to a physical register, this specifies the register that
1015 /// llvm.savestack/llvm.restorestack should save and restore.
1016 void setStackPointerRegisterToSaveRestore(unsigned R) {
1017 StackPointerRegisterToSaveRestore = R;
1020 /// If set to a physical register, this sets the register that receives the
1021 /// exception address on entry to a landing pad.
1022 void setExceptionPointerRegister(unsigned R) {
1023 ExceptionPointerRegister = R;
1026 /// If set to a physical register, this sets the register that receives the
1027 /// exception typeid on entry to a landing pad.
1028 void setExceptionSelectorRegister(unsigned R) {
1029 ExceptionSelectorRegister = R;
1032 /// Tells the code generator not to expand operations into sequences that use
1033 /// the select operations if possible.
1034 void setSelectIsExpensive(bool isExpensive = true) {
1035 SelectIsExpensive = isExpensive;
1038 /// Tells the code generator that the target has multiple (allocatable)
1039 /// condition registers that can be used to store the results of comparisons
1040 /// for use by selects and conditional branches. With multiple condition
1041 /// registers, the code generator will not aggressively sink comparisons into
1042 /// the blocks of their users.
1043 void setHasMultipleConditionRegisters(bool hasManyRegs = true) {
1044 HasMultipleConditionRegisters = hasManyRegs;
1047 /// Tells the code generator that the target has BitExtract instructions.
1048 /// The code generator will aggressively sink "shift"s into the blocks of
1049 /// their users if the users will generate "and" instructions which can be
1050 /// combined with "shift" to BitExtract instructions.
1051 void setHasExtractBitsInsn(bool hasExtractInsn = true) {
1052 HasExtractBitsInsn = hasExtractInsn;
1055 /// Tells the code generator not to expand sequence of operations into a
1056 /// separate sequences that increases the amount of flow control.
1057 void setJumpIsExpensive(bool isExpensive = true) {
1058 JumpIsExpensive = isExpensive;
1061 /// Tells the code generator that integer divide is expensive, and if
1062 /// possible, should be replaced by an alternate sequence of instructions not
1063 /// containing an integer divide.
1064 void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; }
1066 /// Tells the code generator which bitwidths to bypass.
1067 void addBypassSlowDiv(unsigned int SlowBitWidth, unsigned int FastBitWidth) {
1068 BypassSlowDivWidths[SlowBitWidth] = FastBitWidth;
1071 /// Tells the code generator that it shouldn't generate srl/add/sra for a
1072 /// signed divide by power of two, and let the target handle it.
1073 void setPow2DivIsCheap(bool isCheap = true) { Pow2DivIsCheap = isCheap; }
1075 /// Add the specified register class as an available regclass for the
1076 /// specified value type. This indicates the selector can handle values of
1077 /// that class natively.
1078 void addRegisterClass(MVT VT, const TargetRegisterClass *RC) {
1079 assert((unsigned)VT.SimpleTy < array_lengthof(RegClassForVT));
1080 AvailableRegClasses.push_back(std::make_pair(VT, RC));
1081 RegClassForVT[VT.SimpleTy] = RC;
1084 /// Remove all register classes.
1085 void clearRegisterClasses() {
1086 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE * sizeof(TargetRegisterClass*));
1088 AvailableRegClasses.clear();
1091 /// \brief Remove all operation actions.
1092 void clearOperationActions() {
1095 /// Return the largest legal super-reg register class of the register class
1096 /// for the specified type and its associated "cost".
1097 virtual std::pair<const TargetRegisterClass*, uint8_t>
1098 findRepresentativeClass(MVT VT) const;
1100 /// Once all of the register classes are added, this allows us to compute
1101 /// derived properties we expose.
1102 void computeRegisterProperties();
1104 /// Indicate that the specified operation does not work with the specified
1105 /// type and indicate what to do about it.
1106 void setOperationAction(unsigned Op, MVT VT,
1107 LegalizeAction Action) {
1108 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!");
1109 OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action;
1112 /// Indicate that the specified load with extension does not work with the
1113 /// specified type and indicate what to do about it.
1114 void setLoadExtAction(unsigned ExtType, MVT VT,
1115 LegalizeAction Action) {
1116 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE &&
1117 "Table isn't big enough!");
1118 LoadExtActions[VT.SimpleTy][ExtType] = (uint8_t)Action;
1121 /// Indicate that the specified truncating store does not work with the
1122 /// specified type and indicate what to do about it.
1123 void setTruncStoreAction(MVT ValVT, MVT MemVT,
1124 LegalizeAction Action) {
1125 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE &&
1126 "Table isn't big enough!");
1127 TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action;
1130 /// Indicate that the specified indexed load does or does not work with the
1131 /// specified type and indicate what to do abort it.
1133 /// NOTE: All indexed mode loads are initialized to Expand in
1134 /// TargetLowering.cpp
1135 void setIndexedLoadAction(unsigned IdxMode, MVT VT,
1136 LegalizeAction Action) {
1137 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE &&
1138 (unsigned)Action < 0xf && "Table isn't big enough!");
1139 // Load action are kept in the upper half.
1140 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0;
1141 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4;
1144 /// Indicate that the specified indexed store does or does not work with the
1145 /// specified type and indicate what to do about it.
1147 /// NOTE: All indexed mode stores are initialized to Expand in
1148 /// TargetLowering.cpp
1149 void setIndexedStoreAction(unsigned IdxMode, MVT VT,
1150 LegalizeAction Action) {
1151 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE &&
1152 (unsigned)Action < 0xf && "Table isn't big enough!");
1153 // Store action are kept in the lower half.
1154 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f;
1155 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action);
1158 /// Indicate that the specified condition code is or isn't supported on the
1159 /// target and indicate what to do about it.
1160 void setCondCodeAction(ISD::CondCode CC, MVT VT,
1161 LegalizeAction Action) {
1162 assert(VT < MVT::LAST_VALUETYPE &&
1163 (unsigned)CC < array_lengthof(CondCodeActions) &&
1164 "Table isn't big enough!");
1165 /// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 32-bit
1166 /// value and the upper 27 bits index into the second dimension of the array
1167 /// to select what 32-bit value to use.
1168 uint32_t Shift = 2 * (VT.SimpleTy & 0xF);
1169 CondCodeActions[CC][VT.SimpleTy >> 4] &= ~((uint32_t)0x3 << Shift);
1170 CondCodeActions[CC][VT.SimpleTy >> 4] |= (uint32_t)Action << Shift;
1173 /// If Opc/OrigVT is specified as being promoted, the promotion code defaults
1174 /// to trying a larger integer/fp until it can find one that works. If that
1175 /// default is insufficient, this method can be used by the target to override
1177 void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) {
1178 PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy;
1181 /// Targets should invoke this method for each target independent node that
1182 /// they want to provide a custom DAG combiner for by implementing the
1183 /// PerformDAGCombine virtual method.
1184 void setTargetDAGCombine(ISD::NodeType NT) {
1185 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
1186 TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7);
1189 /// Set the target's required jmp_buf buffer size (in bytes); default is 200
1190 void setJumpBufSize(unsigned Size) {
1194 /// Set the target's required jmp_buf buffer alignment (in bytes); default is
1196 void setJumpBufAlignment(unsigned Align) {
1197 JumpBufAlignment = Align;
1200 /// Set the target's minimum function alignment (in log2(bytes))
1201 void setMinFunctionAlignment(unsigned Align) {
1202 MinFunctionAlignment = Align;
1205 /// Set the target's preferred function alignment. This should be set if
1206 /// there is a performance benefit to higher-than-minimum alignment (in
1208 void setPrefFunctionAlignment(unsigned Align) {
1209 PrefFunctionAlignment = Align;
1212 /// Set the target's preferred loop alignment. Default alignment is zero, it
1213 /// means the target does not care about loop alignment. The alignment is
1214 /// specified in log2(bytes).
1215 void setPrefLoopAlignment(unsigned Align) {
1216 PrefLoopAlignment = Align;
1219 /// Set the minimum stack alignment of an argument (in log2(bytes)).
1220 void setMinStackArgumentAlignment(unsigned Align) {
1221 MinStackArgumentAlignment = Align;
1224 /// Set if the DAG builder should automatically insert fences and reduce the
1225 /// order of atomic memory operations to Monotonic.
1226 void setInsertFencesForAtomic(bool fence) {
1227 InsertFencesForAtomic = fence;
1231 //===--------------------------------------------------------------------===//
1232 // Addressing mode description hooks (used by LSR etc).
1235 /// CodeGenPrepare sinks address calculations into the same BB as Load/Store
1236 /// instructions reading the address. This allows as much computation as
1237 /// possible to be done in the address mode for that operand. This hook lets
1238 /// targets also pass back when this should be done on intrinsics which
1240 virtual bool GetAddrModeArguments(IntrinsicInst * /*I*/,
1241 SmallVectorImpl<Value*> &/*Ops*/,
1242 Type *&/*AccessTy*/) const {
1246 /// This represents an addressing mode of:
1247 /// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
1248 /// If BaseGV is null, there is no BaseGV.
1249 /// If BaseOffs is zero, there is no base offset.
1250 /// If HasBaseReg is false, there is no base register.
1251 /// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with
1254 GlobalValue *BaseGV;
1258 AddrMode() : BaseGV(nullptr), BaseOffs(0), HasBaseReg(false), Scale(0) {}
1261 /// Return true if the addressing mode represented by AM is legal for this
1262 /// target, for a load/store of the specified type.
1264 /// The type may be VoidTy, in which case only return true if the addressing
1265 /// mode is legal for a load/store of any legal type. TODO: Handle
1266 /// pre/postinc as well.
1267 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const;
1269 /// \brief Return the cost of the scaling factor used in the addressing mode
1270 /// represented by AM for this target, for a load/store of the specified type.
1272 /// If the AM is supported, the return value must be >= 0.
1273 /// If the AM is not supported, it returns a negative value.
1274 /// TODO: Handle pre/postinc as well.
1275 virtual int getScalingFactorCost(const AddrMode &AM, Type *Ty) const {
1276 // Default: assume that any scaling factor used in a legal AM is free.
1277 if (isLegalAddressingMode(AM, Ty)) return 0;
1281 /// Return true if the specified immediate is legal icmp immediate, that is
1282 /// the target has icmp instructions which can compare a register against the
1283 /// immediate without having to materialize the immediate into a register.
1284 virtual bool isLegalICmpImmediate(int64_t) const {
1288 /// Return true if the specified immediate is legal add immediate, that is the
1289 /// target has add instructions which can add a register with the immediate
1290 /// without having to materialize the immediate into a register.
1291 virtual bool isLegalAddImmediate(int64_t) const {
1295 /// Return true if it's significantly cheaper to shift a vector by a uniform
1296 /// scalar than by an amount which will vary across each lane. On x86, for
1297 /// example, there is a "psllw" instruction for the former case, but no simple
1298 /// instruction for a general "a << b" operation on vectors.
1299 virtual bool isVectorShiftByScalarCheap(Type *Ty) const {
1303 /// Return true if it's free to truncate a value of type Ty1 to type
1304 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
1305 /// by referencing its sub-register AX.
1306 virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const {
1310 /// Return true if a truncation from Ty1 to Ty2 is permitted when deciding
1311 /// whether a call is in tail position. Typically this means that both results
1312 /// would be assigned to the same register or stack slot, but it could mean
1313 /// the target performs adequate checks of its own before proceeding with the
1315 virtual bool allowTruncateForTailCall(Type * /*Ty1*/, Type * /*Ty2*/) const {
1319 virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const {
1323 /// Return true if any actual instruction that defines a value of type Ty1
1324 /// implicitly zero-extends the value to Ty2 in the result register.
1326 /// This does not necessarily include registers defined in unknown ways, such
1327 /// as incoming arguments, or copies from unknown virtual registers. Also, if
1328 /// isTruncateFree(Ty2, Ty1) is true, this does not necessarily apply to
1329 /// truncate instructions. e.g. on x86-64, all instructions that define 32-bit
1330 /// values implicit zero-extend the result out to 64 bits.
1331 virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const {
1335 virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const {
1339 /// Return true if the target supplies and combines to a paired load
1340 /// two loaded values of type LoadedType next to each other in memory.
1341 /// RequiredAlignment gives the minimal alignment constraints that must be met
1342 /// to be able to select this paired load.
1344 /// This information is *not* used to generate actual paired loads, but it is
1345 /// used to generate a sequence of loads that is easier to combine into a
1347 /// For instance, something like this:
1348 /// a = load i64* addr
1349 /// b = trunc i64 a to i32
1350 /// c = lshr i64 a, 32
1351 /// d = trunc i64 c to i32
1352 /// will be optimized into:
1353 /// b = load i32* addr1
1354 /// d = load i32* addr2
1355 /// Where addr1 = addr2 +/- sizeof(i32).
1357 /// In other words, unless the target performs a post-isel load combining,
1358 /// this information should not be provided because it will generate more
1360 virtual bool hasPairedLoad(Type * /*LoadedType*/,
1361 unsigned & /*RequiredAligment*/) const {
1365 virtual bool hasPairedLoad(EVT /*LoadedType*/,
1366 unsigned & /*RequiredAligment*/) const {
1370 /// Return true if zero-extending the specific node Val to type VT2 is free
1371 /// (either because it's implicitly zero-extended such as ARM ldrb / ldrh or
1372 /// because it's folded such as X86 zero-extending loads).
1373 virtual bool isZExtFree(SDValue Val, EVT VT2) const {
1374 return isZExtFree(Val.getValueType(), VT2);
1377 /// Return true if an fneg operation is free to the point where it is never
1378 /// worthwhile to replace it with a bitwise operation.
1379 virtual bool isFNegFree(EVT VT) const {
1380 assert(VT.isFloatingPoint());
1384 /// Return true if an fabs operation is free to the point where it is never
1385 /// worthwhile to replace it with a bitwise operation.
1386 virtual bool isFAbsFree(EVT VT) const {
1387 assert(VT.isFloatingPoint());
1391 /// Return true if an FMA operation is faster than a pair of fmul and fadd
1392 /// instructions. fmuladd intrinsics will be expanded to FMAs when this method
1393 /// returns true, otherwise fmuladd is expanded to fmul + fadd.
1395 /// NOTE: This may be called before legalization on types for which FMAs are
1396 /// not legal, but should return true if those types will eventually legalize
1397 /// to types that support FMAs. After legalization, it will only be called on
1398 /// types that support FMAs (via Legal or Custom actions)
1399 virtual bool isFMAFasterThanFMulAndFAdd(EVT) const {
1403 /// Return true if it's profitable to narrow operations of type VT1 to
1404 /// VT2. e.g. on x86, it's profitable to narrow from i32 to i8 but not from
1406 virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const {
1410 /// \brief Return true if it is beneficial to convert a load of a constant to
1411 /// just the constant itself.
1412 /// On some targets it might be more efficient to use a combination of
1413 /// arithmetic instructions to materialize the constant instead of loading it
1414 /// from a constant pool.
1415 virtual bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
1419 //===--------------------------------------------------------------------===//
1420 // Runtime Library hooks
1423 /// Rename the default libcall routine name for the specified libcall.
1424 void setLibcallName(RTLIB::Libcall Call, const char *Name) {
1425 LibcallRoutineNames[Call] = Name;
1428 /// Get the libcall routine name for the specified libcall.
1429 const char *getLibcallName(RTLIB::Libcall Call) const {
1430 return LibcallRoutineNames[Call];
1433 /// Override the default CondCode to be used to test the result of the
1434 /// comparison libcall against zero.
1435 void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) {
1436 CmpLibcallCCs[Call] = CC;
1439 /// Get the CondCode that's to be used to test the result of the comparison
1440 /// libcall against zero.
1441 ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const {
1442 return CmpLibcallCCs[Call];
1445 /// Set the CallingConv that should be used for the specified libcall.
1446 void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) {
1447 LibcallCallingConvs[Call] = CC;
1450 /// Get the CallingConv that should be used for the specified libcall.
1451 CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const {
1452 return LibcallCallingConvs[Call];
1456 const TargetMachine &TM;
1457 const DataLayout *DL;
1458 const TargetLoweringObjectFile &TLOF;
1460 /// True if this is a little endian target.
1461 bool IsLittleEndian;
1463 /// Tells the code generator not to expand operations into sequences that use
1464 /// the select operations if possible.
1465 bool SelectIsExpensive;
1467 /// Tells the code generator that the target has multiple (allocatable)
1468 /// condition registers that can be used to store the results of comparisons
1469 /// for use by selects and conditional branches. With multiple condition
1470 /// registers, the code generator will not aggressively sink comparisons into
1471 /// the blocks of their users.
1472 bool HasMultipleConditionRegisters;
1474 /// Tells the code generator that the target has BitExtract instructions.
1475 /// The code generator will aggressively sink "shift"s into the blocks of
1476 /// their users if the users will generate "and" instructions which can be
1477 /// combined with "shift" to BitExtract instructions.
1478 bool HasExtractBitsInsn;
1480 /// Tells the code generator not to expand integer divides by constants into a
1481 /// sequence of muls, adds, and shifts. This is a hack until a real cost
1482 /// model is in place. If we ever optimize for size, this will be set to true
1483 /// unconditionally.
1486 /// Tells the code generator to bypass slow divide or remainder
1487 /// instructions. For example, BypassSlowDivWidths[32,8] tells the code
1488 /// generator to bypass 32-bit integer div/rem with an 8-bit unsigned integer
1489 /// div/rem when the operands are positive and less than 256.
1490 DenseMap <unsigned int, unsigned int> BypassSlowDivWidths;
1492 /// Tells the code generator that it shouldn't generate srl/add/sra for a
1493 /// signed divide by power of two, and let the target handle it.
1494 bool Pow2DivIsCheap;
1496 /// Tells the code generator that it shouldn't generate extra flow control
1497 /// instructions and should attempt to combine flow control instructions via
1499 bool JumpIsExpensive;
1501 /// This target prefers to use _setjmp to implement llvm.setjmp.
1503 /// Defaults to false.
1504 bool UseUnderscoreSetJmp;
1506 /// This target prefers to use _longjmp to implement llvm.longjmp.
1508 /// Defaults to false.
1509 bool UseUnderscoreLongJmp;
1511 /// Whether the target can generate code for jumptables. If it's not true,
1512 /// then each jumptable must be lowered into if-then-else's.
1513 bool SupportJumpTables;
1515 /// Number of blocks threshold to use jump tables.
1516 int MinimumJumpTableEntries;
1518 /// Information about the contents of the high-bits in boolean values held in
1519 /// a type wider than i1. See getBooleanContents.
1520 BooleanContent BooleanContents;
1522 /// Information about the contents of the high-bits in boolean values held in
1523 /// a type wider than i1. See getBooleanContents.
1524 BooleanContent BooleanFloatContents;
1526 /// Information about the contents of the high-bits in boolean vector values
1527 /// when the element type is wider than i1. See getBooleanContents.
1528 BooleanContent BooleanVectorContents;
1530 /// The target scheduling preference: shortest possible total cycles or lowest
1532 Sched::Preference SchedPreferenceInfo;
1534 /// The size, in bytes, of the target's jmp_buf buffers
1535 unsigned JumpBufSize;
1537 /// The alignment, in bytes, of the target's jmp_buf buffers
1538 unsigned JumpBufAlignment;
1540 /// The minimum alignment that any argument on the stack needs to have.
1541 unsigned MinStackArgumentAlignment;
1543 /// The minimum function alignment (used when optimizing for size, and to
1544 /// prevent explicitly provided alignment from leading to incorrect code).
1545 unsigned MinFunctionAlignment;
1547 /// The preferred function alignment (used when alignment unspecified and
1548 /// optimizing for speed).
1549 unsigned PrefFunctionAlignment;
1551 /// The preferred loop alignment.
1552 unsigned PrefLoopAlignment;
1554 /// Whether the DAG builder should automatically insert fences and reduce
1555 /// ordering for atomics. (This will be set for for most architectures with
1556 /// weak memory ordering.)
1557 bool InsertFencesForAtomic;
1559 /// If set to a physical register, this specifies the register that
1560 /// llvm.savestack/llvm.restorestack should save and restore.
1561 unsigned StackPointerRegisterToSaveRestore;
1563 /// If set to a physical register, this specifies the register that receives
1564 /// the exception address on entry to a landing pad.
1565 unsigned ExceptionPointerRegister;
1567 /// If set to a physical register, this specifies the register that receives
1568 /// the exception typeid on entry to a landing pad.
1569 unsigned ExceptionSelectorRegister;
1571 /// This indicates the default register class to use for each ValueType the
1572 /// target supports natively.
1573 const TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE];
1574 unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE];
1575 MVT RegisterTypeForVT[MVT::LAST_VALUETYPE];
1577 /// This indicates the "representative" register class to use for each
1578 /// ValueType the target supports natively. This information is used by the
1579 /// scheduler to track register pressure. By default, the representative
1580 /// register class is the largest legal super-reg register class of the
1581 /// register class of the specified type. e.g. On x86, i8, i16, and i32's
1582 /// representative class would be GR32.
1583 const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE];
1585 /// This indicates the "cost" of the "representative" register class for each
1586 /// ValueType. The cost is used by the scheduler to approximate register
1588 uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE];
1590 /// For any value types we are promoting or expanding, this contains the value
1591 /// type that we are changing to. For Expanded types, this contains one step
1592 /// of the expand (e.g. i64 -> i32), even if there are multiple steps required
1593 /// (e.g. i64 -> i16). For types natively supported by the system, this holds
1594 /// the same type (e.g. i32 -> i32).
1595 MVT TransformToType[MVT::LAST_VALUETYPE];
1597 /// For each operation and each value type, keep a LegalizeAction that
1598 /// indicates how instruction selection should deal with the operation. Most
1599 /// operations are Legal (aka, supported natively by the target), but
1600 /// operations that are not should be described. Note that operations on
1601 /// non-legal value types are not described here.
1602 uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END];
1604 /// For each load extension type and each value type, keep a LegalizeAction
1605 /// that indicates how instruction selection should deal with a load of a
1606 /// specific value type and extension type.
1607 uint8_t LoadExtActions[MVT::LAST_VALUETYPE][ISD::LAST_LOADEXT_TYPE];
1609 /// For each value type pair keep a LegalizeAction that indicates whether a
1610 /// truncating store of a specific value type and truncating type is legal.
1611 uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE];
1613 /// For each indexed mode and each value type, keep a pair of LegalizeAction
1614 /// that indicates how instruction selection should deal with the load /
1617 /// The first dimension is the value_type for the reference. The second
1618 /// dimension represents the various modes for load store.
1619 uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE];
1621 /// For each condition code (ISD::CondCode) keep a LegalizeAction that
1622 /// indicates how instruction selection should deal with the condition code.
1624 /// Because each CC action takes up 2 bits, we need to have the array size be
1625 /// large enough to fit all of the value types. This can be done by rounding
1626 /// up the MVT::LAST_VALUETYPE value to the next multiple of 16.
1627 uint32_t CondCodeActions[ISD::SETCC_INVALID][(MVT::LAST_VALUETYPE + 15) / 16];
1629 ValueTypeActionImpl ValueTypeActions;
1633 getTypeConversion(LLVMContext &Context, EVT VT) const {
1634 // If this is a simple type, use the ComputeRegisterProp mechanism.
1635 if (VT.isSimple()) {
1636 MVT SVT = VT.getSimpleVT();
1637 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
1638 MVT NVT = TransformToType[SVT.SimpleTy];
1639 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
1642 (LA == TypeLegal || LA == TypeSoftenFloat ||
1643 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)
1644 && "Promote may not follow Expand or Promote");
1646 if (LA == TypeSplitVector)
1647 return LegalizeKind(LA, EVT::getVectorVT(Context,
1648 SVT.getVectorElementType(),
1649 SVT.getVectorNumElements()/2));
1650 if (LA == TypeScalarizeVector)
1651 return LegalizeKind(LA, SVT.getVectorElementType());
1652 return LegalizeKind(LA, NVT);
1655 // Handle Extended Scalar Types.
1656 if (!VT.isVector()) {
1657 assert(VT.isInteger() && "Float types must be simple");
1658 unsigned BitSize = VT.getSizeInBits();
1659 // First promote to a power-of-two size, then expand if necessary.
1660 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
1661 EVT NVT = VT.getRoundIntegerType(Context);
1662 assert(NVT != VT && "Unable to round integer VT");
1663 LegalizeKind NextStep = getTypeConversion(Context, NVT);
1664 // Avoid multi-step promotion.
1665 if (NextStep.first == TypePromoteInteger) return NextStep;
1666 // Return rounded integer type.
1667 return LegalizeKind(TypePromoteInteger, NVT);
1670 return LegalizeKind(TypeExpandInteger,
1671 EVT::getIntegerVT(Context, VT.getSizeInBits()/2));
1674 // Handle vector types.
1675 unsigned NumElts = VT.getVectorNumElements();
1676 EVT EltVT = VT.getVectorElementType();
1678 // Vectors with only one element are always scalarized.
1680 return LegalizeKind(TypeScalarizeVector, EltVT);
1682 // Try to widen vector elements until the element type is a power of two and
1683 // promote it to a legal type later on, for example:
1684 // <3 x i8> -> <4 x i8> -> <4 x i32>
1685 if (EltVT.isInteger()) {
1686 // Vectors with a number of elements that is not a power of two are always
1687 // widened, for example <3 x i8> -> <4 x i8>.
1688 if (!VT.isPow2VectorType()) {
1689 NumElts = (unsigned)NextPowerOf2(NumElts);
1690 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
1691 return LegalizeKind(TypeWidenVector, NVT);
1694 // Examine the element type.
1695 LegalizeKind LK = getTypeConversion(Context, EltVT);
1697 // If type is to be expanded, split the vector.
1698 // <4 x i140> -> <2 x i140>
1699 if (LK.first == TypeExpandInteger)
1700 return LegalizeKind(TypeSplitVector,
1701 EVT::getVectorVT(Context, EltVT, NumElts / 2));
1703 // Promote the integer element types until a legal vector type is found
1704 // or until the element integer type is too big. If a legal type was not
1705 // found, fallback to the usual mechanism of widening/splitting the
1707 EVT OldEltVT = EltVT;
1709 // Increase the bitwidth of the element to the next pow-of-two
1710 // (which is greater than 8 bits).
1711 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()
1712 ).getRoundIntegerType(Context);
1714 // Stop trying when getting a non-simple element type.
1715 // Note that vector elements may be greater than legal vector element
1716 // types. Example: X86 XMM registers hold 64bit element on 32bit
1718 if (!EltVT.isSimple()) break;
1720 // Build a new vector type and check if it is legal.
1721 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1722 // Found a legal promoted vector type.
1723 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
1724 return LegalizeKind(TypePromoteInteger,
1725 EVT::getVectorVT(Context, EltVT, NumElts));
1728 // Reset the type to the unexpanded type if we did not find a legal vector
1729 // type with a promoted vector element type.
1733 // Try to widen the vector until a legal type is found.
1734 // If there is no wider legal type, split the vector.
1736 // Round up to the next power of 2.
1737 NumElts = (unsigned)NextPowerOf2(NumElts);
1739 // If there is no simple vector type with this many elements then there
1740 // cannot be a larger legal vector type. Note that this assumes that
1741 // there are no skipped intermediate vector types in the simple types.
1742 if (!EltVT.isSimple()) break;
1743 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1744 if (LargerVector == MVT()) break;
1746 // If this type is legal then widen the vector.
1747 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
1748 return LegalizeKind(TypeWidenVector, LargerVector);
1751 // Widen odd vectors to next power of two.
1752 if (!VT.isPow2VectorType()) {
1753 EVT NVT = VT.getPow2VectorType(Context);
1754 return LegalizeKind(TypeWidenVector, NVT);
1757 // Vectors with illegal element types are expanded.
1758 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
1759 return LegalizeKind(TypeSplitVector, NVT);
1763 std::vector<std::pair<MVT, const TargetRegisterClass*> > AvailableRegClasses;
1765 /// Targets can specify ISD nodes that they would like PerformDAGCombine
1766 /// callbacks for by calling setTargetDAGCombine(), which sets a bit in this
1769 TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT];
1771 /// For operations that must be promoted to a specific type, this holds the
1772 /// destination type. This map should be sparse, so don't hold it as an
1775 /// Targets add entries to this map with AddPromotedToType(..), clients access
1776 /// this with getTypeToPromoteTo(..).
1777 std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType>
1780 /// Stores the name each libcall.
1781 const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL];
1783 /// The ISD::CondCode that should be used to test the result of each of the
1784 /// comparison libcall against zero.
1785 ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL];
1787 /// Stores the CallingConv that should be used for each libcall.
1788 CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL];
1791 /// \brief Specify maximum number of store instructions per memset call.
1793 /// When lowering \@llvm.memset this field specifies the maximum number of
1794 /// store operations that may be substituted for the call to memset. Targets
1795 /// must set this value based on the cost threshold for that target. Targets
1796 /// should assume that the memset will be done using as many of the largest
1797 /// store operations first, followed by smaller ones, if necessary, per
1798 /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine
1799 /// with 16-bit alignment would result in four 2-byte stores and one 1-byte
1800 /// store. This only applies to setting a constant array of a constant size.
1801 unsigned MaxStoresPerMemset;
1803 /// Maximum number of stores operations that may be substituted for the call
1804 /// to memset, used for functions with OptSize attribute.
1805 unsigned MaxStoresPerMemsetOptSize;
1807 /// \brief Specify maximum bytes of store instructions per memcpy call.
1809 /// When lowering \@llvm.memcpy this field specifies the maximum number of
1810 /// store operations that may be substituted for a call to memcpy. Targets
1811 /// must set this value based on the cost threshold for that target. Targets
1812 /// should assume that the memcpy will be done using as many of the largest
1813 /// store operations first, followed by smaller ones, if necessary, per
1814 /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine
1815 /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store
1816 /// and one 1-byte store. This only applies to copying a constant array of
1818 unsigned MaxStoresPerMemcpy;
1820 /// Maximum number of store operations that may be substituted for a call to
1821 /// memcpy, used for functions with OptSize attribute.
1822 unsigned MaxStoresPerMemcpyOptSize;
1824 /// \brief Specify maximum bytes of store instructions per memmove call.
1826 /// When lowering \@llvm.memmove this field specifies the maximum number of
1827 /// store instructions that may be substituted for a call to memmove. Targets
1828 /// must set this value based on the cost threshold for that target. Targets
1829 /// should assume that the memmove will be done using as many of the largest
1830 /// store operations first, followed by smaller ones, if necessary, per
1831 /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine
1832 /// with 8-bit alignment would result in nine 1-byte stores. This only
1833 /// applies to copying a constant array of constant size.
1834 unsigned MaxStoresPerMemmove;
1836 /// Maximum number of store instructions that may be substituted for a call to
1837 /// memmove, used for functions with OpSize attribute.
1838 unsigned MaxStoresPerMemmoveOptSize;
1840 /// Tells the code generator that select is more expensive than a branch if
1841 /// the branch is usually predicted right.
1842 bool PredictableSelectIsExpensive;
1844 /// MaskAndBranchFoldingIsLegal - Indicates if the target supports folding
1845 /// a mask of a single bit, a compare, and a branch into a single instruction.
1846 bool MaskAndBranchFoldingIsLegal;
1849 /// Return true if the value types that can be represented by the specified
1850 /// register class are all legal.
1851 bool isLegalRC(const TargetRegisterClass *RC) const;
1853 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
1854 /// sequence of memory operands that is recognized by PrologEpilogInserter.
1855 MachineBasicBlock *emitPatchPoint(MachineInstr *MI, MachineBasicBlock *MBB) const;
1858 /// This class defines information used to lower LLVM code to legal SelectionDAG
1859 /// operators that the target instruction selector can accept natively.
1861 /// This class also defines callbacks that targets must implement to lower
1862 /// target-specific constructs to SelectionDAG operators.
1863 class TargetLowering : public TargetLoweringBase {
1864 TargetLowering(const TargetLowering&) LLVM_DELETED_FUNCTION;
1865 void operator=(const TargetLowering&) LLVM_DELETED_FUNCTION;
1868 /// NOTE: The constructor takes ownership of TLOF.
1869 explicit TargetLowering(const TargetMachine &TM,
1870 const TargetLoweringObjectFile *TLOF);
1872 /// Returns true by value, base pointer and offset pointer and addressing mode
1873 /// by reference if the node's address can be legally represented as
1874 /// pre-indexed load / store address.
1875 virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/,
1876 SDValue &/*Offset*/,
1877 ISD::MemIndexedMode &/*AM*/,
1878 SelectionDAG &/*DAG*/) const {
1882 /// Returns true by value, base pointer and offset pointer and addressing mode
1883 /// by reference if this node can be combined with a load / store to form a
1884 /// post-indexed load / store.
1885 virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/,
1887 SDValue &/*Offset*/,
1888 ISD::MemIndexedMode &/*AM*/,
1889 SelectionDAG &/*DAG*/) const {
1893 /// Return the entry encoding for a jump table in the current function. The
1894 /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum.
1895 virtual unsigned getJumpTableEncoding() const;
1897 virtual const MCExpr *
1898 LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/,
1899 const MachineBasicBlock * /*MBB*/, unsigned /*uid*/,
1900 MCContext &/*Ctx*/) const {
1901 llvm_unreachable("Need to implement this hook if target has custom JTIs");
1904 /// Returns relocation base for the given PIC jumptable.
1905 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
1906 SelectionDAG &DAG) const;
1908 /// This returns the relocation base for the given PIC jumptable, the same as
1909 /// getPICJumpTableRelocBase, but as an MCExpr.
1910 virtual const MCExpr *
1911 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1912 unsigned JTI, MCContext &Ctx) const;
1914 /// Return true if folding a constant offset with the given GlobalAddress is
1915 /// legal. It is frequently not legal in PIC relocation models.
1916 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
1918 bool isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
1919 SDValue &Chain) const;
1921 void softenSetCCOperands(SelectionDAG &DAG, EVT VT,
1922 SDValue &NewLHS, SDValue &NewRHS,
1923 ISD::CondCode &CCCode, SDLoc DL) const;
1925 /// Returns a pair of (return value, chain).
1926 std::pair<SDValue, SDValue> makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC,
1927 EVT RetVT, const SDValue *Ops,
1928 unsigned NumOps, bool isSigned,
1929 SDLoc dl, bool doesNotReturn = false,
1930 bool isReturnValueUsed = true) const;
1932 //===--------------------------------------------------------------------===//
1933 // TargetLowering Optimization Methods
1936 /// A convenience struct that encapsulates a DAG, and two SDValues for
1937 /// returning information from TargetLowering to its clients that want to
1939 struct TargetLoweringOpt {
1946 explicit TargetLoweringOpt(SelectionDAG &InDAG,
1948 DAG(InDAG), LegalTys(LT), LegalOps(LO) {}
1950 bool LegalTypes() const { return LegalTys; }
1951 bool LegalOperations() const { return LegalOps; }
1953 bool CombineTo(SDValue O, SDValue N) {
1959 /// Check to see if the specified operand of the specified instruction is a
1960 /// constant integer. If so, check to see if there are any bits set in the
1961 /// constant that are not demanded. If so, shrink the constant and return
1963 bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded);
1965 /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free. This
1966 /// uses isZExtFree and ZERO_EXTEND for the widening cast, but it could be
1967 /// generalized for targets with other types of implicit widening casts.
1968 bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded,
1972 /// Look at Op. At this point, we know that only the DemandedMask bits of the
1973 /// result of Op are ever used downstream. If we can use this information to
1974 /// simplify Op, create a new simplified DAG node and return true, returning
1975 /// the original and new nodes in Old and New. Otherwise, analyze the
1976 /// expression and return a mask of KnownOne and KnownZero bits for the
1977 /// expression (used to simplify the caller). The KnownZero/One bits may only
1978 /// be accurate for those bits in the DemandedMask.
1979 bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask,
1980 APInt &KnownZero, APInt &KnownOne,
1981 TargetLoweringOpt &TLO, unsigned Depth = 0) const;
1983 /// Determine which of the bits specified in Mask are known to be either zero
1984 /// or one and return them in the KnownZero/KnownOne bitsets.
1985 virtual void computeKnownBitsForTargetNode(const SDValue Op,
1988 const SelectionDAG &DAG,
1989 unsigned Depth = 0) const;
1991 /// This method can be implemented by targets that want to expose additional
1992 /// information about sign bits to the DAG Combiner.
1993 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
1994 const SelectionDAG &DAG,
1995 unsigned Depth = 0) const;
1997 struct DAGCombinerInfo {
1998 void *DC; // The DAG Combiner object.
2000 bool CalledByLegalizer;
2004 DAGCombinerInfo(SelectionDAG &dag, CombineLevel level, bool cl, void *dc)
2005 : DC(dc), Level(level), CalledByLegalizer(cl), DAG(dag) {}
2007 bool isBeforeLegalize() const { return Level == BeforeLegalizeTypes; }
2008 bool isBeforeLegalizeOps() const { return Level < AfterLegalizeVectorOps; }
2009 bool isAfterLegalizeVectorOps() const {
2010 return Level == AfterLegalizeDAG;
2012 CombineLevel getDAGCombineLevel() { return Level; }
2013 bool isCalledByLegalizer() const { return CalledByLegalizer; }
2015 void AddToWorklist(SDNode *N);
2016 void RemoveFromWorklist(SDNode *N);
2017 SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To,
2019 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true);
2020 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true);
2022 void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO);
2025 /// Return if the N is a constant or constant vector equal to the true value
2026 /// from getBooleanContents().
2027 bool isConstTrueVal(const SDNode *N) const;
2029 /// Return if the N is a constant or constant vector equal to the false value
2030 /// from getBooleanContents().
2031 bool isConstFalseVal(const SDNode *N) const;
2033 /// Try to simplify a setcc built with the specified operands and cc. If it is
2034 /// unable to simplify it, return a null SDValue.
2035 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
2036 ISD::CondCode Cond, bool foldBooleans,
2037 DAGCombinerInfo &DCI, SDLoc dl) const;
2039 /// Returns true (and the GlobalValue and the offset) if the node is a
2040 /// GlobalAddress + offset.
2042 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
2044 /// This method will be invoked for all target nodes and for any
2045 /// target-independent nodes that the target has registered with invoke it
2048 /// The semantics are as follows:
2050 /// SDValue.Val == 0 - No change was made
2051 /// SDValue.Val == N - N was replaced, is dead, and is already handled.
2052 /// otherwise - N should be replaced by the returned Operand.
2054 /// In addition, methods provided by DAGCombinerInfo may be used to perform
2055 /// more complex transformations.
2057 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
2059 /// Return true if it is profitable to move a following shift through this
2060 // node, adjusting any immediate operands as necessary to preserve semantics.
2061 // This transformation may not be desirable if it disrupts a particularly
2062 // auspicious target-specific tree (e.g. bitfield extraction in AArch64).
2063 // By default, it returns true.
2064 virtual bool isDesirableToCommuteWithShift(const SDNode *N /*Op*/) const {
2068 /// Return true if the target has native support for the specified value type
2069 /// and it is 'desirable' to use the type for the given node type. e.g. On x86
2070 /// i16 is legal, but undesirable since i16 instruction encodings are longer
2071 /// and some i16 instructions are slow.
2072 virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const {
2073 // By default, assume all legal types are desirable.
2074 return isTypeLegal(VT);
2077 /// Return true if it is profitable for dag combiner to transform a floating
2078 /// point op of specified opcode to a equivalent op of an integer
2079 /// type. e.g. f32 load -> i32 load can be profitable on ARM.
2080 virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/,
2085 /// This method query the target whether it is beneficial for dag combiner to
2086 /// promote the specified node. If true, it should return the desired
2087 /// promotion type by reference.
2088 virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const {
2092 //===--------------------------------------------------------------------===//
2093 // Lowering methods - These methods must be implemented by targets so that
2094 // the SelectionDAGBuilder code knows how to lower these.
2097 /// This hook must be implemented to lower the incoming (formal) arguments,
2098 /// described by the Ins array, into the specified DAG. The implementation
2099 /// should fill in the InVals array with legal-type argument values, and
2100 /// return the resulting token chain value.
2103 LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/,
2105 const SmallVectorImpl<ISD::InputArg> &/*Ins*/,
2106 SDLoc /*dl*/, SelectionDAG &/*DAG*/,
2107 SmallVectorImpl<SDValue> &/*InVals*/) const {
2108 llvm_unreachable("Not Implemented");
2111 struct ArgListEntry {
2120 bool isInAlloca : 1;
2121 bool isReturned : 1;
2124 ArgListEntry() : isSExt(false), isZExt(false), isInReg(false),
2125 isSRet(false), isNest(false), isByVal(false), isInAlloca(false),
2126 isReturned(false), Alignment(0) { }
2128 void setAttributes(ImmutableCallSite *CS, unsigned AttrIdx);
2130 typedef std::vector<ArgListEntry> ArgListTy;
2132 /// This structure contains all information that is necessary for lowering
2133 /// calls. It is passed to TLI::LowerCallTo when the SelectionDAG builder
2134 /// needs to lower a call, and targets will see this struct in their LowerCall
2136 struct CallLoweringInfo {
2143 bool DoesNotReturn : 1;
2144 bool IsReturnValueUsed : 1;
2146 // IsTailCall should be modified by implementations of
2147 // TargetLowering::LowerCall that perform tail call conversions.
2150 unsigned NumFixedArgs;
2151 CallingConv::ID CallConv;
2156 ImmutableCallSite *CS;
2157 SmallVector<ISD::OutputArg, 32> Outs;
2158 SmallVector<SDValue, 32> OutVals;
2159 SmallVector<ISD::InputArg, 32> Ins;
2161 CallLoweringInfo(SelectionDAG &DAG)
2162 : RetTy(nullptr), RetSExt(false), RetZExt(false), IsVarArg(false),
2163 IsInReg(false), DoesNotReturn(false), IsReturnValueUsed(true),
2164 IsTailCall(false), NumFixedArgs(-1), CallConv(CallingConv::C),
2165 DAG(DAG), CS(nullptr) {}
2167 CallLoweringInfo &setDebugLoc(SDLoc dl) {
2172 CallLoweringInfo &setChain(SDValue InChain) {
2177 CallLoweringInfo &setCallee(CallingConv::ID CC, Type *ResultType,
2178 SDValue Target, ArgListTy &&ArgsList,
2179 unsigned FixedArgs = -1) {
2184 (FixedArgs == static_cast<unsigned>(-1) ? Args.size() : FixedArgs);
2185 Args = std::move(ArgsList);
2189 CallLoweringInfo &setCallee(Type *ResultType, FunctionType *FTy,
2190 SDValue Target, ArgListTy &&ArgsList,
2191 ImmutableCallSite &Call) {
2194 IsInReg = Call.paramHasAttr(0, Attribute::InReg);
2195 DoesNotReturn = Call.doesNotReturn();
2196 IsVarArg = FTy->isVarArg();
2197 IsReturnValueUsed = !Call.getInstruction()->use_empty();
2198 RetSExt = Call.paramHasAttr(0, Attribute::SExt);
2199 RetZExt = Call.paramHasAttr(0, Attribute::ZExt);
2203 CallConv = Call.getCallingConv();
2204 NumFixedArgs = FTy->getNumParams();
2205 Args = std::move(ArgsList);
2212 CallLoweringInfo &setInRegister(bool Value = true) {
2217 CallLoweringInfo &setNoReturn(bool Value = true) {
2218 DoesNotReturn = Value;
2222 CallLoweringInfo &setVarArg(bool Value = true) {
2227 CallLoweringInfo &setTailCall(bool Value = true) {
2232 CallLoweringInfo &setDiscardResult(bool Value = true) {
2233 IsReturnValueUsed = !Value;
2237 CallLoweringInfo &setSExtResult(bool Value = true) {
2242 CallLoweringInfo &setZExtResult(bool Value = true) {
2247 ArgListTy &getArgs() {
2252 /// This function lowers an abstract call to a function into an actual call.
2253 /// This returns a pair of operands. The first element is the return value
2254 /// for the function (if RetTy is not VoidTy). The second element is the
2255 /// outgoing token chain. It calls LowerCall to do the actual lowering.
2256 std::pair<SDValue, SDValue> LowerCallTo(CallLoweringInfo &CLI) const;
2258 /// This hook must be implemented to lower calls into the the specified
2259 /// DAG. The outgoing arguments to the call are described by the Outs array,
2260 /// and the values to be returned by the call are described by the Ins
2261 /// array. The implementation should fill in the InVals array with legal-type
2262 /// return values from the call, and return the resulting token chain value.
2264 LowerCall(CallLoweringInfo &/*CLI*/,
2265 SmallVectorImpl<SDValue> &/*InVals*/) const {
2266 llvm_unreachable("Not Implemented");
2269 /// Target-specific cleanup for formal ByVal parameters.
2270 virtual void HandleByVal(CCState *, unsigned &, unsigned) const {}
2272 /// This hook should be implemented to check whether the return values
2273 /// described by the Outs array can fit into the return registers. If false
2274 /// is returned, an sret-demotion is performed.
2275 virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/,
2276 MachineFunction &/*MF*/, bool /*isVarArg*/,
2277 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/,
2278 LLVMContext &/*Context*/) const
2280 // Return true by default to get preexisting behavior.
2284 /// This hook must be implemented to lower outgoing return values, described
2285 /// by the Outs array, into the specified DAG. The implementation should
2286 /// return the resulting token chain value.
2288 LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/,
2290 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/,
2291 const SmallVectorImpl<SDValue> &/*OutVals*/,
2292 SDLoc /*dl*/, SelectionDAG &/*DAG*/) const {
2293 llvm_unreachable("Not Implemented");
2296 /// Return true if result of the specified node is used by a return node
2297 /// only. It also compute and return the input chain for the tail call.
2299 /// This is used to determine whether it is possible to codegen a libcall as
2300 /// tail call at legalization time.
2301 virtual bool isUsedByReturnOnly(SDNode *, SDValue &/*Chain*/) const {
2305 /// Return true if the target may be able emit the call instruction as a tail
2306 /// call. This is used by optimization passes to determine if it's profitable
2307 /// to duplicate return instructions to enable tailcall optimization.
2308 virtual bool mayBeEmittedAsTailCall(CallInst *) const {
2312 /// Return the builtin name for the __builtin___clear_cache intrinsic
2313 /// Default is to invoke the clear cache library call
2314 virtual const char * getClearCacheBuiltinName() const {
2315 return "__clear_cache";
2318 /// Return the register ID of the name passed in. Used by named register
2319 /// global variables extension. There is no target-independent behaviour
2320 /// so the default action is to bail.
2321 virtual unsigned getRegisterByName(const char* RegName, EVT VT) const {
2322 report_fatal_error("Named registers not implemented for this target");
2325 /// Return the type that should be used to zero or sign extend a
2326 /// zeroext/signext integer argument or return value. FIXME: Most C calling
2327 /// convention requires the return type to be promoted, but this is not true
2328 /// all the time, e.g. i1 on x86-64. It is also not necessary for non-C
2329 /// calling conventions. The frontend should handle this and include all of
2330 /// the necessary information.
2331 virtual MVT getTypeForExtArgOrReturn(MVT VT,
2332 ISD::NodeType /*ExtendKind*/) const {
2333 MVT MinVT = getRegisterType(MVT::i32);
2334 return VT.bitsLT(MinVT) ? MinVT : VT;
2337 /// For some targets, an LLVM struct type must be broken down into multiple
2338 /// simple types, but the calling convention specifies that the entire struct
2339 /// must be passed in a block of consecutive registers.
2341 functionArgumentNeedsConsecutiveRegisters(Type *Ty, CallingConv::ID CallConv,
2342 bool isVarArg) const {
2346 /// Returns a 0 terminated array of registers that can be safely used as
2347 /// scratch registers.
2348 virtual const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const {
2352 /// This callback is used to prepare for a volatile or atomic load.
2353 /// It takes a chain node as input and returns the chain for the load itself.
2355 /// Having a callback like this is necessary for targets like SystemZ,
2356 /// which allows a CPU to reuse the result of a previous load indefinitely,
2357 /// even if a cache-coherent store is performed by another CPU. The default
2358 /// implementation does nothing.
2359 virtual SDValue prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL,
2360 SelectionDAG &DAG) const {
2364 /// This callback is invoked by the type legalizer to legalize nodes with an
2365 /// illegal operand type but legal result types. It replaces the
2366 /// LowerOperation callback in the type Legalizer. The reason we can not do
2367 /// away with LowerOperation entirely is that LegalizeDAG isn't yet ready to
2368 /// use this callback.
2370 /// TODO: Consider merging with ReplaceNodeResults.
2372 /// The target places new result values for the node in Results (their number
2373 /// and types must exactly match those of the original return values of
2374 /// the node), or leaves Results empty, which indicates that the node is not
2375 /// to be custom lowered after all.
2376 /// The default implementation calls LowerOperation.
2377 virtual void LowerOperationWrapper(SDNode *N,
2378 SmallVectorImpl<SDValue> &Results,
2379 SelectionDAG &DAG) const;
2381 /// This callback is invoked for operations that are unsupported by the
2382 /// target, which are registered to use 'custom' lowering, and whose defined
2383 /// values are all legal. If the target has no operations that require custom
2384 /// lowering, it need not implement this. The default implementation of this
2386 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
2388 /// This callback is invoked when a node result type is illegal for the
2389 /// target, and the operation was registered to use 'custom' lowering for that
2390 /// result type. The target places new result values for the node in Results
2391 /// (their number and types must exactly match those of the original return
2392 /// values of the node), or leaves Results empty, which indicates that the
2393 /// node is not to be custom lowered after all.
2395 /// If the target has no operations that require custom lowering, it need not
2396 /// implement this. The default implementation aborts.
2397 virtual void ReplaceNodeResults(SDNode * /*N*/,
2398 SmallVectorImpl<SDValue> &/*Results*/,
2399 SelectionDAG &/*DAG*/) const {
2400 llvm_unreachable("ReplaceNodeResults not implemented for this target!");
2403 /// This method returns the name of a target specific DAG node.
2404 virtual const char *getTargetNodeName(unsigned Opcode) const;
2406 /// This method returns a target specific FastISel object, or null if the
2407 /// target does not support "fast" ISel.
2408 virtual FastISel *createFastISel(FunctionLoweringInfo &,
2409 const TargetLibraryInfo *) const {
2414 bool verifyReturnAddressArgumentIsConstant(SDValue Op,
2415 SelectionDAG &DAG) const;
2417 //===--------------------------------------------------------------------===//
2418 // Inline Asm Support hooks
2421 /// This hook allows the target to expand an inline asm call to be explicit
2422 /// llvm code if it wants to. This is useful for turning simple inline asms
2423 /// into LLVM intrinsics, which gives the compiler more information about the
2424 /// behavior of the code.
2425 virtual bool ExpandInlineAsm(CallInst *) const {
2429 enum ConstraintType {
2430 C_Register, // Constraint represents specific register(s).
2431 C_RegisterClass, // Constraint represents any of register(s) in class.
2432 C_Memory, // Memory constraint.
2433 C_Other, // Something else.
2434 C_Unknown // Unsupported constraint.
2437 enum ConstraintWeight {
2439 CW_Invalid = -1, // No match.
2440 CW_Okay = 0, // Acceptable.
2441 CW_Good = 1, // Good weight.
2442 CW_Better = 2, // Better weight.
2443 CW_Best = 3, // Best weight.
2445 // Well-known weights.
2446 CW_SpecificReg = CW_Okay, // Specific register operands.
2447 CW_Register = CW_Good, // Register operands.
2448 CW_Memory = CW_Better, // Memory operands.
2449 CW_Constant = CW_Best, // Constant operand.
2450 CW_Default = CW_Okay // Default or don't know type.
2453 /// This contains information for each constraint that we are lowering.
2454 struct AsmOperandInfo : public InlineAsm::ConstraintInfo {
2455 /// This contains the actual string for the code, like "m". TargetLowering
2456 /// picks the 'best' code from ConstraintInfo::Codes that most closely
2457 /// matches the operand.
2458 std::string ConstraintCode;
2460 /// Information about the constraint code, e.g. Register, RegisterClass,
2461 /// Memory, Other, Unknown.
2462 TargetLowering::ConstraintType ConstraintType;
2464 /// If this is the result output operand or a clobber, this is null,
2465 /// otherwise it is the incoming operand to the CallInst. This gets
2466 /// modified as the asm is processed.
2467 Value *CallOperandVal;
2469 /// The ValueType for the operand value.
2472 /// Return true of this is an input operand that is a matching constraint
2474 bool isMatchingInputConstraint() const;
2476 /// If this is an input matching constraint, this method returns the output
2477 /// operand it matches.
2478 unsigned getMatchedOperand() const;
2480 /// Copy constructor for copying from a ConstraintInfo.
2481 AsmOperandInfo(const InlineAsm::ConstraintInfo &info)
2482 : InlineAsm::ConstraintInfo(info),
2483 ConstraintType(TargetLowering::C_Unknown),
2484 CallOperandVal(nullptr), ConstraintVT(MVT::Other) {
2488 typedef std::vector<AsmOperandInfo> AsmOperandInfoVector;
2490 /// Split up the constraint string from the inline assembly value into the
2491 /// specific constraints and their prefixes, and also tie in the associated
2492 /// operand values. If this returns an empty vector, and if the constraint
2493 /// string itself isn't empty, there was an error parsing.
2494 virtual AsmOperandInfoVector ParseConstraints(ImmutableCallSite CS) const;
2496 /// Examine constraint type and operand type and determine a weight value.
2497 /// The operand object must already have been set up with the operand type.
2498 virtual ConstraintWeight getMultipleConstraintMatchWeight(
2499 AsmOperandInfo &info, int maIndex) const;
2501 /// Examine constraint string and operand type and determine a weight value.
2502 /// The operand object must already have been set up with the operand type.
2503 virtual ConstraintWeight getSingleConstraintMatchWeight(
2504 AsmOperandInfo &info, const char *constraint) const;
2506 /// Determines the constraint code and constraint type to use for the specific
2507 /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType.
2508 /// If the actual operand being passed in is available, it can be passed in as
2509 /// Op, otherwise an empty SDValue can be passed.
2510 virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2512 SelectionDAG *DAG = nullptr) const;
2514 /// Given a constraint, return the type of constraint it is for this target.
2515 virtual ConstraintType getConstraintType(const std::string &Constraint) const;
2517 /// Given a physical register constraint (e.g. {edx}), return the register
2518 /// number and the register class for the register.
2520 /// Given a register class constraint, like 'r', if this corresponds directly
2521 /// to an LLVM register class, return a register of 0 and the register class
2524 /// This should only be used for C_Register constraints. On error, this
2525 /// returns a register number of 0 and a null register class pointer..
2526 virtual std::pair<unsigned, const TargetRegisterClass*>
2527 getRegForInlineAsmConstraint(const std::string &Constraint,
2530 /// Try to replace an X constraint, which matches anything, with another that
2531 /// has more specific requirements based on the type of the corresponding
2532 /// operand. This returns null if there is no replacement to make.
2533 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
2535 /// Lower the specified operand into the Ops vector. If it is invalid, don't
2536 /// add anything to Ops.
2537 virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
2538 std::vector<SDValue> &Ops,
2539 SelectionDAG &DAG) const;
2541 //===--------------------------------------------------------------------===//
2542 // Div utility functions
2544 SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl,
2545 SelectionDAG &DAG) const;
2546 SDValue BuildSDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
2547 bool IsAfterLegalization,
2548 std::vector<SDNode *> *Created) const;
2549 SDValue BuildUDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
2550 bool IsAfterLegalization,
2551 std::vector<SDNode *> *Created) const;
2552 virtual SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor,
2554 std::vector<SDNode *> *Created) const {
2558 //===--------------------------------------------------------------------===//
2559 // Legalization utility functions
2562 /// Expand a MUL into two nodes. One that computes the high bits of
2563 /// the result and one that computes the low bits.
2564 /// \param HiLoVT The value type to use for the Lo and Hi nodes.
2565 /// \param LL Low bits of the LHS of the MUL. You can use this parameter
2566 /// if you want to control how low bits are extracted from the LHS.
2567 /// \param LH High bits of the LHS of the MUL. See LL for meaning.
2568 /// \param RL Low bits of the RHS of the MUL. See LL for meaning
2569 /// \param RH High bits of the RHS of the MUL. See LL for meaning.
2570 /// \returns true if the node has been expanded. false if it has not
2571 bool expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT,
2572 SelectionDAG &DAG, SDValue LL = SDValue(),
2573 SDValue LH = SDValue(), SDValue RL = SDValue(),
2574 SDValue RH = SDValue()) const;
2576 /// Expand float(f32) to SINT(i64) conversion
2577 /// \param N Node to expand
2578 /// \param Result output after conversion
2579 /// \returns True, if the expansion was successful, false otherwise
2580 bool expandFP_TO_SINT(SDNode *N, SDValue &Result, SelectionDAG &DAG) const;
2582 //===--------------------------------------------------------------------===//
2583 // Instruction Emitting Hooks
2586 /// This method should be implemented by targets that mark instructions with
2587 /// the 'usesCustomInserter' flag. These instructions are special in various
2588 /// ways, which require special support to insert. The specified MachineInstr
2589 /// is created but not inserted into any basic blocks, and this method is
2590 /// called to expand it into a sequence of instructions, potentially also
2591 /// creating new basic blocks and control flow.
2592 virtual MachineBasicBlock *
2593 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const;
2595 /// This method should be implemented by targets that mark instructions with
2596 /// the 'hasPostISelHook' flag. These instructions must be adjusted after
2597 /// instruction selection by target hooks. e.g. To fill in optional defs for
2598 /// ARM 's' setting instructions.
2600 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const;
2602 /// If this function returns true, SelectionDAGBuilder emits a
2603 /// LOAD_STACK_GUARD node when it is lowering Intrinsic::stackprotector.
2604 virtual bool useLoadStackGuardNode() const {
2609 /// Given an LLVM IR type and return type attributes, compute the return value
2610 /// EVTs and flags, and optionally also the offsets, if the return value is
2611 /// being lowered to memory.
2612 void GetReturnInfo(Type* ReturnType, AttributeSet attr,
2613 SmallVectorImpl<ISD::OutputArg> &Outs,
2614 const TargetLowering &TLI);
2616 } // end llvm namespace