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 //===----------------------------------------------------------------------===//
10 // This file describes how to lower LLVM code to machine code. This has two
13 // 1. Which ValueTypes are natively supported by the target.
14 // 2. Which operations are supported for supported ValueTypes.
15 // 3. Cost thresholds for alternative implementations of certain operations.
17 // In addition it has a few other components, like information about FP
20 //===----------------------------------------------------------------------===//
22 #ifndef LLVM_TARGET_TARGETLOWERING_H
23 #define LLVM_TARGET_TARGETLOWERING_H
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/CodeGen/DAGCombine.h"
27 #include "llvm/CodeGen/RuntimeLibcalls.h"
28 #include "llvm/CodeGen/SelectionDAGNodes.h"
29 #include "llvm/IR/Attributes.h"
30 #include "llvm/IR/CallingConv.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/Support/CallSite.h"
33 #include "llvm/Support/DebugLoc.h"
34 #include "llvm/Target/TargetCallingConv.h"
35 #include "llvm/Target/TargetMachine.h"
44 class FunctionLoweringInfo;
45 class ImmutableCallSite;
47 class MachineBasicBlock;
48 class MachineFunction;
50 class MachineJumpTableInfo;
53 template<typename T> class SmallVectorImpl;
55 class TargetRegisterClass;
56 class TargetLibraryInfo;
57 class TargetLoweringObjectFile;
62 None, // No preference
63 Source, // Follow source order.
64 RegPressure, // Scheduling for lowest register pressure.
65 Hybrid, // Scheduling for both latency and register pressure.
66 ILP, // Scheduling for ILP in low register pressure mode.
67 VLIW // Scheduling for VLIW targets.
71 /// TargetLoweringBase - This base class for TargetLowering contains the
72 /// SelectionDAG-independent parts that can be used from the rest of CodeGen.
73 class TargetLoweringBase {
74 TargetLoweringBase(const TargetLoweringBase&) LLVM_DELETED_FUNCTION;
75 void operator=(const TargetLoweringBase&) LLVM_DELETED_FUNCTION;
78 /// LegalizeAction - This enum indicates whether operations are valid for a
79 /// target, and if not, what action should be used to make them valid.
81 Legal, // The target natively supports this operation.
82 Promote, // This operation should be executed in a larger type.
83 Expand, // Try to expand this to other ops, otherwise use a libcall.
84 Custom // Use the LowerOperation hook to implement custom lowering.
87 /// LegalizeTypeAction - This enum indicates whether a types are legal for a
88 /// target, and if not, what action should be used to make them valid.
89 enum LegalizeTypeAction {
90 TypeLegal, // The target natively supports this type.
91 TypePromoteInteger, // Replace this integer with a larger one.
92 TypeExpandInteger, // Split this integer into two of half the size.
93 TypeSoftenFloat, // Convert this float to a same size integer type.
94 TypeExpandFloat, // Split this float into two of half the size.
95 TypeScalarizeVector, // Replace this one-element vector with its element.
96 TypeSplitVector, // Split this vector into two of half the size.
97 TypeWidenVector // This vector should be widened into a larger vector.
100 /// LegalizeKind holds the legalization kind that needs to happen to EVT
101 /// in order to type-legalize it.
102 typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind;
104 enum BooleanContent { // How the target represents true/false values.
105 UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage.
106 ZeroOrOneBooleanContent, // All bits zero except for bit 0.
107 ZeroOrNegativeOneBooleanContent // All bits equal to bit 0.
110 enum SelectSupportKind {
111 ScalarValSelect, // The target supports scalar selects (ex: cmov).
112 ScalarCondVectorVal, // The target supports selects with a scalar condition
113 // and vector values (ex: cmov).
114 VectorMaskSelect // The target supports vector selects with a vector
115 // mask (ex: x86 blends).
118 static ISD::NodeType getExtendForContent(BooleanContent Content) {
120 case UndefinedBooleanContent:
121 // Extend by adding rubbish bits.
122 return ISD::ANY_EXTEND;
123 case ZeroOrOneBooleanContent:
124 // Extend by adding zero bits.
125 return ISD::ZERO_EXTEND;
126 case ZeroOrNegativeOneBooleanContent:
127 // Extend by copying the sign bit.
128 return ISD::SIGN_EXTEND;
130 llvm_unreachable("Invalid content kind");
133 /// NOTE: The constructor takes ownership of TLOF.
134 explicit TargetLoweringBase(const TargetMachine &TM,
135 const TargetLoweringObjectFile *TLOF);
136 virtual ~TargetLoweringBase();
139 /// \brief Initialize all of the actions to default values.
143 const TargetMachine &getTargetMachine() const { return TM; }
144 const DataLayout *getDataLayout() const { return TD; }
145 const TargetLoweringObjectFile &getObjFileLowering() const { return TLOF; }
147 bool isBigEndian() const { return !IsLittleEndian; }
148 bool isLittleEndian() const { return IsLittleEndian; }
149 // Return the pointer type for the given address space, defaults to
150 // the pointer type from the data layout.
151 // FIXME: The default needs to be removed once all the code is updated.
152 virtual MVT getPointerTy(uint32_t AS = 0) const { return PointerTy; }
153 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const;
155 EVT getShiftAmountTy(EVT LHSTy) const;
157 /// isSelectExpensive - Return true if the select operation is expensive for
159 bool isSelectExpensive() const { return SelectIsExpensive; }
161 virtual bool isSelectSupported(SelectSupportKind kind) const { return true; }
163 /// shouldSplitVectorElementType - Return true if a vector of the given type
164 /// should be split (TypeSplitVector) instead of promoted
165 /// (TypePromoteInteger) during type legalization.
166 virtual bool shouldSplitVectorElementType(EVT VT) const { return false; }
168 /// isIntDivCheap() - Return true if integer divide is usually cheaper than
169 /// a sequence of several shifts, adds, and multiplies for this target.
170 bool isIntDivCheap() const { return IntDivIsCheap; }
172 /// isSlowDivBypassed - Returns true if target has indicated at least one
173 /// type should be bypassed.
174 bool isSlowDivBypassed() const { return !BypassSlowDivWidths.empty(); }
176 /// getBypassSlowDivTypes - Returns map of slow types for division or
177 /// remainder with corresponding fast types
178 const DenseMap<unsigned int, unsigned int> &getBypassSlowDivWidths() const {
179 return BypassSlowDivWidths;
182 /// isPow2DivCheap() - Return true if pow2 div is cheaper than a chain of
184 bool isPow2DivCheap() const { return Pow2DivIsCheap; }
186 /// isJumpExpensive() - Return true if Flow Control is an expensive operation
187 /// that should be avoided.
188 bool isJumpExpensive() const { return JumpIsExpensive; }
190 /// isPredictableSelectExpensive - Return true if selects are only cheaper
191 /// than branches if the branch is unlikely to be predicted right.
192 bool isPredictableSelectExpensive() const {
193 return PredictableSelectIsExpensive;
196 /// getSetCCResultType - Return the ValueType of the result of SETCC
197 /// operations. Also used to obtain the target's preferred type for
198 /// the condition operand of SELECT and BRCOND nodes. In the case of
199 /// BRCOND the argument passed is MVT::Other since there are no other
200 /// operands to get a type hint from.
201 virtual EVT getSetCCResultType(EVT VT) const;
203 /// getCmpLibcallReturnType - Return the ValueType for comparison
204 /// libcalls. Comparions libcalls include floating point comparion calls,
205 /// and Ordered/Unordered check calls on floating point numbers.
207 MVT::SimpleValueType getCmpLibcallReturnType() const;
209 /// getBooleanContents - For targets without i1 registers, this gives the
210 /// nature of the high-bits of boolean values held in types wider than i1.
211 /// "Boolean values" are special true/false values produced by nodes like
212 /// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND.
213 /// Not to be confused with general values promoted from i1.
214 /// Some cpus distinguish between vectors of boolean and scalars; the isVec
215 /// parameter selects between the two kinds. For example on X86 a scalar
216 /// boolean should be zero extended from i1, while the elements of a vector
217 /// of booleans should be sign extended from i1.
218 BooleanContent getBooleanContents(bool isVec) const {
219 return isVec ? BooleanVectorContents : BooleanContents;
222 /// getSchedulingPreference - Return target scheduling preference.
223 Sched::Preference getSchedulingPreference() const {
224 return SchedPreferenceInfo;
227 /// getSchedulingPreference - Some scheduler, e.g. hybrid, can switch to
228 /// different scheduling heuristics for different nodes. This function returns
229 /// the preference (or none) for the given node.
230 virtual Sched::Preference getSchedulingPreference(SDNode *) const {
234 /// getRegClassFor - Return the register class that should be used for the
235 /// specified value type.
236 virtual const TargetRegisterClass *getRegClassFor(MVT VT) const {
237 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
238 assert(RC && "This value type is not natively supported!");
242 /// getRepRegClassFor - Return the 'representative' register class for the
243 /// specified value type. The 'representative' register class is the largest
244 /// legal super-reg register class for the register class of the value type.
245 /// For example, on i386 the rep register class for i8, i16, and i32 are GR32;
246 /// while the rep register class is GR64 on x86_64.
247 virtual const TargetRegisterClass *getRepRegClassFor(MVT VT) const {
248 const TargetRegisterClass *RC = RepRegClassForVT[VT.SimpleTy];
252 /// getRepRegClassCostFor - Return the cost of the 'representative' register
253 /// class for the specified value type.
254 virtual uint8_t getRepRegClassCostFor(MVT VT) const {
255 return RepRegClassCostForVT[VT.SimpleTy];
258 /// isTypeLegal - Return true if the target has native support for the
259 /// specified value type. This means that it has a register that directly
260 /// holds it without promotions or expansions.
261 bool isTypeLegal(EVT VT) const {
262 assert(!VT.isSimple() ||
263 (unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT));
264 return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != 0;
267 class ValueTypeActionImpl {
268 /// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum
269 /// that indicates how instruction selection should deal with the type.
270 uint8_t ValueTypeActions[MVT::LAST_VALUETYPE];
273 ValueTypeActionImpl() {
274 std::fill(ValueTypeActions, array_endof(ValueTypeActions), 0);
277 LegalizeTypeAction getTypeAction(MVT VT) const {
278 return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy];
281 void setTypeAction(MVT VT, LegalizeTypeAction Action) {
282 unsigned I = VT.SimpleTy;
283 ValueTypeActions[I] = Action;
287 const ValueTypeActionImpl &getValueTypeActions() const {
288 return ValueTypeActions;
291 /// getTypeAction - Return how we should legalize values of this type, either
292 /// it is already legal (return 'Legal') or we need to promote it to a larger
293 /// type (return 'Promote'), or we need to expand it into multiple registers
294 /// of smaller integer type (return 'Expand'). 'Custom' is not an option.
295 LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const {
296 return getTypeConversion(Context, VT).first;
298 LegalizeTypeAction getTypeAction(MVT VT) const {
299 return ValueTypeActions.getTypeAction(VT);
302 /// getTypeToTransformTo - For types supported by the target, this is an
303 /// identity function. For types that must be promoted to larger types, this
304 /// returns the larger type to promote to. For integer types that are larger
305 /// than the largest integer register, this contains one step in the expansion
306 /// to get to the smaller register. For illegal floating point types, this
307 /// returns the integer type to transform to.
308 EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const {
309 return getTypeConversion(Context, VT).second;
312 /// getTypeToExpandTo - For types supported by the target, this is an
313 /// identity function. For types that must be expanded (i.e. integer types
314 /// that are larger than the largest integer register or illegal floating
315 /// point types), this returns the largest legal type it will be expanded to.
316 EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const {
317 assert(!VT.isVector());
319 switch (getTypeAction(Context, VT)) {
322 case TypeExpandInteger:
323 VT = getTypeToTransformTo(Context, VT);
326 llvm_unreachable("Type is not legal nor is it to be expanded!");
331 /// getVectorTypeBreakdown - Vector types are broken down into some number of
332 /// legal first class types. For example, EVT::v8f32 maps to 2 EVT::v4f32
333 /// with Altivec or SSE1, or 8 promoted EVT::f64 values with the X86 FP stack.
334 /// Similarly, EVT::v2i64 turns into 4 EVT::i32 values with both PPC and X86.
336 /// This method returns the number of registers needed, and the VT for each
337 /// register. It also returns the VT and quantity of the intermediate values
338 /// before they are promoted/expanded.
340 unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
342 unsigned &NumIntermediates,
343 MVT &RegisterVT) const;
345 /// getTgtMemIntrinsic: Given an intrinsic, checks if on the target the
346 /// intrinsic will need to map to a MemIntrinsicNode (touches memory). If
347 /// this is the case, it returns true and store the intrinsic
348 /// information into the IntrinsicInfo that was passed to the function.
349 struct IntrinsicInfo {
350 unsigned opc; // target opcode
351 EVT memVT; // memory VT
352 const Value* ptrVal; // value representing memory location
353 int offset; // offset off of ptrVal
354 unsigned align; // alignment
355 bool vol; // is volatile?
356 bool readMem; // reads memory?
357 bool writeMem; // writes memory?
360 virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &,
361 unsigned /*Intrinsic*/) const {
365 /// isFPImmLegal - Returns true if the target can instruction select the
366 /// specified FP immediate natively. If false, the legalizer will materialize
367 /// the FP immediate as a load from a constant pool.
368 virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const {
372 /// isShuffleMaskLegal - Targets can use this to indicate that they only
373 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
374 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
375 /// are assumed to be legal.
376 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/,
381 /// canOpTrap - Returns true if the operation can trap for the value type.
382 /// VT must be a legal type. By default, we optimistically assume most
383 /// operations don't trap except for divide and remainder.
384 virtual bool canOpTrap(unsigned Op, EVT VT) const;
386 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
387 /// used by Targets can use this to indicate if there is a suitable
388 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
390 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/,
395 /// getOperationAction - Return how this operation should be treated: either
396 /// it is legal, needs to be promoted to a larger size, needs to be
397 /// expanded to some other code sequence, or the target has a custom expander
399 LegalizeAction getOperationAction(unsigned Op, EVT VT) const {
400 if (VT.isExtended()) return Expand;
401 // If a target-specific SDNode requires legalization, require the target
402 // to provide custom legalization for it.
403 if (Op > array_lengthof(OpActions[0])) return Custom;
404 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy;
405 return (LegalizeAction)OpActions[I][Op];
408 /// isOperationLegalOrCustom - Return true if the specified operation is
409 /// legal on this target or can be made legal with custom lowering. This
410 /// is used to help guide high-level lowering decisions.
411 bool isOperationLegalOrCustom(unsigned Op, EVT VT) const {
412 return (VT == MVT::Other || isTypeLegal(VT)) &&
413 (getOperationAction(Op, VT) == Legal ||
414 getOperationAction(Op, VT) == Custom);
417 /// isOperationLegalOrPromote - Return true if the specified operation is
418 /// legal on this target or can be made legal using promotion. This
419 /// is used to help guide high-level lowering decisions.
420 bool isOperationLegalOrPromote(unsigned Op, EVT VT) const {
421 return (VT == MVT::Other || isTypeLegal(VT)) &&
422 (getOperationAction(Op, VT) == Legal ||
423 getOperationAction(Op, VT) == Promote);
426 /// isOperationExpand - Return true if the specified operation is illegal on
427 /// this target or unlikely to be made legal with custom lowering. This is
428 /// used to help guide high-level lowering decisions.
429 bool isOperationExpand(unsigned Op, EVT VT) const {
430 return (!isTypeLegal(VT) || getOperationAction(Op, VT) == Expand);
433 /// isOperationLegal - Return true if the specified operation is legal on this
435 bool isOperationLegal(unsigned Op, EVT VT) const {
436 return (VT == MVT::Other || isTypeLegal(VT)) &&
437 getOperationAction(Op, VT) == Legal;
440 /// getLoadExtAction - Return how this load with extension should be treated:
441 /// either it is legal, needs to be promoted to a larger size, needs to be
442 /// expanded to some other code sequence, or the target has a custom expander
444 LegalizeAction getLoadExtAction(unsigned ExtType, MVT VT) const {
445 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE &&
446 "Table isn't big enough!");
447 return (LegalizeAction)LoadExtActions[VT.SimpleTy][ExtType];
450 /// isLoadExtLegal - Return true if the specified load with extension is legal
452 bool isLoadExtLegal(unsigned ExtType, EVT VT) const {
453 return VT.isSimple() &&
454 getLoadExtAction(ExtType, VT.getSimpleVT()) == Legal;
457 /// getTruncStoreAction - Return how this store with truncation should be
458 /// treated: either it is legal, needs to be promoted to a larger size, needs
459 /// to be expanded to some other code sequence, or the target has a custom
461 LegalizeAction getTruncStoreAction(MVT ValVT, MVT MemVT) const {
462 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE &&
463 "Table isn't big enough!");
464 return (LegalizeAction)TruncStoreActions[ValVT.SimpleTy]
468 /// isTruncStoreLegal - Return true if the specified store with truncation is
469 /// legal on this target.
470 bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const {
471 return isTypeLegal(ValVT) && MemVT.isSimple() &&
472 getTruncStoreAction(ValVT.getSimpleVT(), MemVT.getSimpleVT()) == Legal;
475 /// getIndexedLoadAction - Return how the indexed load should be treated:
476 /// either it is legal, needs to be promoted to a larger size, needs to be
477 /// expanded to some other code sequence, or the target has a custom expander
480 getIndexedLoadAction(unsigned IdxMode, MVT VT) const {
481 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE &&
482 "Table isn't big enough!");
483 unsigned Ty = (unsigned)VT.SimpleTy;
484 return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4);
487 /// isIndexedLoadLegal - Return true if the specified indexed load is legal
489 bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const {
490 return VT.isSimple() &&
491 (getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Legal ||
492 getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Custom);
495 /// getIndexedStoreAction - Return how the indexed store should be treated:
496 /// either it is legal, needs to be promoted to a larger size, needs to be
497 /// expanded to some other code sequence, or the target has a custom expander
500 getIndexedStoreAction(unsigned IdxMode, MVT VT) const {
501 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE &&
502 "Table isn't big enough!");
503 unsigned Ty = (unsigned)VT.SimpleTy;
504 return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f);
507 /// isIndexedStoreLegal - Return true if the specified indexed load is legal
509 bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const {
510 return VT.isSimple() &&
511 (getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Legal ||
512 getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Custom);
515 /// getCondCodeAction - Return how the condition code should be treated:
516 /// either it is legal, needs to be expanded to some other code sequence,
517 /// or the target has a custom expander for it.
519 getCondCodeAction(ISD::CondCode CC, MVT VT) const {
520 assert((unsigned)CC < array_lengthof(CondCodeActions) &&
521 (unsigned)VT.SimpleTy < sizeof(CondCodeActions[0])*4 &&
522 "Table isn't big enough!");
523 /// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 64bit
524 /// value and the upper 27 bits index into the second dimension of the
525 /// array to select what 64bit value to use.
526 LegalizeAction Action = (LegalizeAction)
527 ((CondCodeActions[CC][VT.SimpleTy >> 5] >> (2*(VT.SimpleTy & 0x1F))) & 3);
528 assert(Action != Promote && "Can't promote condition code!");
532 /// isCondCodeLegal - Return true if the specified condition code is legal
534 bool isCondCodeLegal(ISD::CondCode CC, MVT VT) const {
536 getCondCodeAction(CC, VT) == Legal ||
537 getCondCodeAction(CC, VT) == Custom;
541 /// getTypeToPromoteTo - If the action for this operation is to promote, this
542 /// method returns the ValueType to promote to.
543 MVT getTypeToPromoteTo(unsigned Op, MVT VT) const {
544 assert(getOperationAction(Op, VT) == Promote &&
545 "This operation isn't promoted!");
547 // See if this has an explicit type specified.
548 std::map<std::pair<unsigned, MVT::SimpleValueType>,
549 MVT::SimpleValueType>::const_iterator PTTI =
550 PromoteToType.find(std::make_pair(Op, VT.SimpleTy));
551 if (PTTI != PromoteToType.end()) return PTTI->second;
553 assert((VT.isInteger() || VT.isFloatingPoint()) &&
554 "Cannot autopromote this type, add it with AddPromotedToType.");
558 NVT = (MVT::SimpleValueType)(NVT.SimpleTy+1);
559 assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid &&
560 "Didn't find type to promote to!");
561 } while (!isTypeLegal(NVT) ||
562 getOperationAction(Op, NVT) == Promote);
566 /// getValueType - Return the EVT corresponding to this LLVM type.
567 /// This is fixed by the LLVM operations except for the pointer size. If
568 /// AllowUnknown is true, this will return MVT::Other for types with no EVT
569 /// counterpart (e.g. structs), otherwise it will assert.
570 EVT getValueType(Type *Ty, bool AllowUnknown = false) const {
571 // Lower scalar pointers to native pointer types.
572 if (Ty->isPointerTy()) return PointerTy;
574 if (Ty->isVectorTy()) {
575 VectorType *VTy = cast<VectorType>(Ty);
576 Type *Elm = VTy->getElementType();
577 // Lower vectors of pointers to native pointer types.
578 if (Elm->isPointerTy())
579 Elm = EVT(PointerTy).getTypeForEVT(Ty->getContext());
580 return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false),
581 VTy->getNumElements());
583 return EVT::getEVT(Ty, AllowUnknown);
586 /// Return the MVT corresponding to this LLVM type. See getValueType.
587 MVT getSimpleValueType(Type *Ty, bool AllowUnknown = false) const {
588 return getValueType(Ty, AllowUnknown).getSimpleVT();
591 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
592 /// function arguments in the caller parameter area. This is the actual
593 /// alignment, not its logarithm.
594 virtual unsigned getByValTypeAlignment(Type *Ty) const;
596 /// getRegisterType - Return the type of registers that this ValueType will
597 /// eventually require.
598 MVT getRegisterType(MVT VT) const {
599 assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT));
600 return RegisterTypeForVT[VT.SimpleTy];
603 /// getRegisterType - Return the type of registers that this ValueType will
604 /// eventually require.
605 MVT getRegisterType(LLVMContext &Context, EVT VT) const {
607 assert((unsigned)VT.getSimpleVT().SimpleTy <
608 array_lengthof(RegisterTypeForVT));
609 return RegisterTypeForVT[VT.getSimpleVT().SimpleTy];
614 unsigned NumIntermediates;
615 (void)getVectorTypeBreakdown(Context, VT, VT1,
616 NumIntermediates, RegisterVT);
619 if (VT.isInteger()) {
620 return getRegisterType(Context, getTypeToTransformTo(Context, VT));
622 llvm_unreachable("Unsupported extended type!");
625 /// getNumRegisters - Return the number of registers that this ValueType will
626 /// eventually require. This is one for any types promoted to live in larger
627 /// registers, but may be more than one for types (like i64) that are split
628 /// into pieces. For types like i140, which are first promoted then expanded,
629 /// it is the number of registers needed to hold all the bits of the original
630 /// type. For an i140 on a 32 bit machine this means 5 registers.
631 unsigned getNumRegisters(LLVMContext &Context, EVT VT) const {
633 assert((unsigned)VT.getSimpleVT().SimpleTy <
634 array_lengthof(NumRegistersForVT));
635 return NumRegistersForVT[VT.getSimpleVT().SimpleTy];
640 unsigned NumIntermediates;
641 return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2);
643 if (VT.isInteger()) {
644 unsigned BitWidth = VT.getSizeInBits();
645 unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits();
646 return (BitWidth + RegWidth - 1) / RegWidth;
648 llvm_unreachable("Unsupported extended type!");
651 /// ShouldShrinkFPConstant - If true, then instruction selection should
652 /// seek to shrink the FP constant of the specified type to a smaller type
653 /// in order to save space and / or reduce runtime.
654 virtual bool ShouldShrinkFPConstant(EVT) const { return true; }
656 /// hasTargetDAGCombine - If true, the target has custom DAG combine
657 /// transformations that it can perform for the specified node.
658 bool hasTargetDAGCombine(ISD::NodeType NT) const {
659 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
660 return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7));
663 /// This function returns the maximum number of store operations permitted
664 /// to replace a call to llvm.memset. The value is set by the target at the
665 /// performance threshold for such a replacement. If OptSize is true,
666 /// return the limit for functions that have OptSize attribute.
667 /// @brief Get maximum # of store operations permitted for llvm.memset
668 unsigned getMaxStoresPerMemset(bool OptSize) const {
669 return OptSize ? MaxStoresPerMemsetOptSize : MaxStoresPerMemset;
672 /// This function returns the maximum number of store operations permitted
673 /// to replace a call to llvm.memcpy. The value is set by the target at the
674 /// performance threshold for such a replacement. If OptSize is true,
675 /// return the limit for functions that have OptSize attribute.
676 /// @brief Get maximum # of store operations permitted for llvm.memcpy
677 unsigned getMaxStoresPerMemcpy(bool OptSize) const {
678 return OptSize ? MaxStoresPerMemcpyOptSize : MaxStoresPerMemcpy;
681 /// This function returns the maximum number of store operations permitted
682 /// to replace a call to llvm.memmove. The value is set by the target at the
683 /// performance threshold for such a replacement. If OptSize is true,
684 /// return the limit for functions that have OptSize attribute.
685 /// @brief Get maximum # of store operations permitted for llvm.memmove
686 unsigned getMaxStoresPerMemmove(bool OptSize) const {
687 return OptSize ? MaxStoresPerMemmoveOptSize : MaxStoresPerMemmove;
690 /// This function returns true if the target allows unaligned memory accesses.
691 /// of the specified type. If true, it also returns whether the unaligned
692 /// memory access is "fast" in the second argument by reference. This is used,
693 /// for example, in situations where an array copy/move/set is converted to a
694 /// sequence of store operations. It's use helps to ensure that such
695 /// replacements don't generate code that causes an alignment error (trap) on
696 /// the target machine.
697 /// @brief Determine if the target supports unaligned memory accesses.
698 virtual bool allowsUnalignedMemoryAccesses(EVT, bool *Fast = 0) const {
702 /// getOptimalMemOpType - Returns the target specific optimal type for load
703 /// and store operations as a result of memset, memcpy, and memmove
704 /// lowering. If DstAlign is zero that means it's safe to destination
705 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
706 /// means there isn't a need to check it against alignment requirement,
707 /// probably because the source does not need to be loaded. If 'IsMemset' is
708 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
709 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
710 /// source is constant so it does not need to be loaded.
711 /// It returns EVT::Other if the type should be determined using generic
712 /// target-independent logic.
713 virtual EVT getOptimalMemOpType(uint64_t /*Size*/,
714 unsigned /*DstAlign*/, unsigned /*SrcAlign*/,
717 bool /*MemcpyStrSrc*/,
718 MachineFunction &/*MF*/) const {
722 /// isSafeMemOpType - Returns true if it's safe to use load / store of the
723 /// specified type to expand memcpy / memset inline. This is mostly true
724 /// for all types except for some special cases. For example, on X86
725 /// targets without SSE2 f64 load / store are done with fldl / fstpl which
726 /// also does type conversion. Note the specified type doesn't have to be
727 /// legal as the hook is used before type legalization.
728 virtual bool isSafeMemOpType(MVT VT) const {
732 /// usesUnderscoreSetJmp - Determine if we should use _setjmp or setjmp
733 /// to implement llvm.setjmp.
734 bool usesUnderscoreSetJmp() const {
735 return UseUnderscoreSetJmp;
738 /// usesUnderscoreLongJmp - Determine if we should use _longjmp or longjmp
739 /// to implement llvm.longjmp.
740 bool usesUnderscoreLongJmp() const {
741 return UseUnderscoreLongJmp;
744 /// supportJumpTables - return whether the target can generate code for
746 bool supportJumpTables() const {
747 return SupportJumpTables;
750 /// getMinimumJumpTableEntries - return integer threshold on number of
751 /// blocks to use jump tables rather than if sequence.
752 int getMinimumJumpTableEntries() const {
753 return MinimumJumpTableEntries;
756 /// getStackPointerRegisterToSaveRestore - If a physical register, this
757 /// specifies the register that llvm.savestack/llvm.restorestack should save
759 unsigned getStackPointerRegisterToSaveRestore() const {
760 return StackPointerRegisterToSaveRestore;
763 /// getExceptionPointerRegister - If a physical register, this returns
764 /// the register that receives the exception address on entry to a landing
766 unsigned getExceptionPointerRegister() const {
767 return ExceptionPointerRegister;
770 /// getExceptionSelectorRegister - If a physical register, this returns
771 /// the register that receives the exception typeid on entry to a landing
773 unsigned getExceptionSelectorRegister() const {
774 return ExceptionSelectorRegister;
777 /// getJumpBufSize - returns the target's jmp_buf size in bytes (if never
778 /// set, the default is 200)
779 unsigned getJumpBufSize() const {
783 /// getJumpBufAlignment - returns the target's jmp_buf alignment in bytes
784 /// (if never set, the default is 0)
785 unsigned getJumpBufAlignment() const {
786 return JumpBufAlignment;
789 /// getMinStackArgumentAlignment - return the minimum stack alignment of an
791 unsigned getMinStackArgumentAlignment() const {
792 return MinStackArgumentAlignment;
795 /// getMinFunctionAlignment - return the minimum function alignment.
797 unsigned getMinFunctionAlignment() const {
798 return MinFunctionAlignment;
801 /// getPrefFunctionAlignment - return the preferred function alignment.
803 unsigned getPrefFunctionAlignment() const {
804 return PrefFunctionAlignment;
807 /// getPrefLoopAlignment - return the preferred loop alignment.
809 unsigned getPrefLoopAlignment() const {
810 return PrefLoopAlignment;
813 /// getShouldFoldAtomicFences - return whether the combiner should fold
814 /// fence MEMBARRIER instructions into the atomic intrinsic instructions.
816 bool getShouldFoldAtomicFences() const {
817 return ShouldFoldAtomicFences;
820 /// getInsertFencesFor - return whether the DAG builder should automatically
821 /// insert fences and reduce ordering for atomics.
823 bool getInsertFencesForAtomic() const {
824 return InsertFencesForAtomic;
827 /// getStackCookieLocation - Return true if the target stores stack
828 /// protector cookies at a fixed offset in some non-standard address
829 /// space, and populates the address space and offset as
831 virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/,
832 unsigned &/*Offset*/) const {
836 /// getMaximalGlobalOffset - Returns the maximal possible offset which can be
837 /// used for loads / stores from the global.
838 virtual unsigned getMaximalGlobalOffset() const {
842 //===--------------------------------------------------------------------===//
843 /// \name Helpers for TargetTransformInfo implementations
846 /// Get the ISD node that corresponds to the Instruction class opcode.
847 int InstructionOpcodeToISD(unsigned Opcode) const;
849 /// Estimate the cost of type-legalization and the legalized type.
850 std::pair<unsigned, MVT> getTypeLegalizationCost(Type *Ty) const;
854 //===--------------------------------------------------------------------===//
855 // TargetLowering Configuration Methods - These methods should be invoked by
856 // the derived class constructor to configure this object for the target.
859 /// \brief Reset the operation actions based on target options.
860 virtual void resetOperationActions() {}
863 /// setBooleanContents - Specify how the target extends the result of a
864 /// boolean value from i1 to a wider type. See getBooleanContents.
865 void setBooleanContents(BooleanContent Ty) { BooleanContents = Ty; }
866 /// setBooleanVectorContents - Specify how the target extends the result
867 /// of a vector boolean value from a vector of i1 to a wider type. See
868 /// getBooleanContents.
869 void setBooleanVectorContents(BooleanContent Ty) {
870 BooleanVectorContents = Ty;
873 /// setSchedulingPreference - Specify the target scheduling preference.
874 void setSchedulingPreference(Sched::Preference Pref) {
875 SchedPreferenceInfo = Pref;
878 /// setUseUnderscoreSetJmp - Indicate whether this target prefers to
879 /// use _setjmp to implement llvm.setjmp or the non _ version.
880 /// Defaults to false.
881 void setUseUnderscoreSetJmp(bool Val) {
882 UseUnderscoreSetJmp = Val;
885 /// setUseUnderscoreLongJmp - Indicate whether this target prefers to
886 /// use _longjmp to implement llvm.longjmp or the non _ version.
887 /// Defaults to false.
888 void setUseUnderscoreLongJmp(bool Val) {
889 UseUnderscoreLongJmp = Val;
892 /// setSupportJumpTables - Indicate whether the target can generate code for
894 void setSupportJumpTables(bool Val) {
895 SupportJumpTables = Val;
898 /// setMinimumJumpTableEntries - Indicate the number of blocks to generate
899 /// jump tables rather than if sequence.
900 void setMinimumJumpTableEntries(int Val) {
901 MinimumJumpTableEntries = Val;
904 /// setStackPointerRegisterToSaveRestore - If set to a physical register, this
905 /// specifies the register that llvm.savestack/llvm.restorestack should save
907 void setStackPointerRegisterToSaveRestore(unsigned R) {
908 StackPointerRegisterToSaveRestore = R;
911 /// setExceptionPointerRegister - If set to a physical register, this sets
912 /// the register that receives the exception address on entry to a landing
914 void setExceptionPointerRegister(unsigned R) {
915 ExceptionPointerRegister = R;
918 /// setExceptionSelectorRegister - If set to a physical register, this sets
919 /// the register that receives the exception typeid on entry to a landing
921 void setExceptionSelectorRegister(unsigned R) {
922 ExceptionSelectorRegister = R;
925 /// SelectIsExpensive - Tells the code generator not to expand operations
926 /// into sequences that use the select operations if possible.
927 void setSelectIsExpensive(bool isExpensive = true) {
928 SelectIsExpensive = isExpensive;
931 /// JumpIsExpensive - Tells the code generator not to expand sequence of
932 /// operations into a separate sequences that increases the amount of
934 void setJumpIsExpensive(bool isExpensive = true) {
935 JumpIsExpensive = isExpensive;
938 /// setIntDivIsCheap - Tells the code generator that integer divide is
939 /// expensive, and if possible, should be replaced by an alternate sequence
940 /// of instructions not containing an integer divide.
941 void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; }
943 /// addBypassSlowDiv - Tells the code generator which bitwidths to bypass.
944 void addBypassSlowDiv(unsigned int SlowBitWidth, unsigned int FastBitWidth) {
945 BypassSlowDivWidths[SlowBitWidth] = FastBitWidth;
948 /// setPow2DivIsCheap - Tells the code generator that it shouldn't generate
949 /// srl/add/sra for a signed divide by power of two, and let the target handle
951 void setPow2DivIsCheap(bool isCheap = true) { Pow2DivIsCheap = isCheap; }
953 /// addRegisterClass - Add the specified register class as an available
954 /// regclass for the specified value type. This indicates the selector can
955 /// handle values of that class natively.
956 void addRegisterClass(MVT VT, const TargetRegisterClass *RC) {
957 assert((unsigned)VT.SimpleTy < array_lengthof(RegClassForVT));
958 AvailableRegClasses.push_back(std::make_pair(VT, RC));
959 RegClassForVT[VT.SimpleTy] = RC;
962 /// clearRegisterClasses - Remove all register classes.
963 void clearRegisterClasses() {
964 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE * sizeof(TargetRegisterClass*));
966 AvailableRegClasses.clear();
969 /// \brief Remove all operation actions.
970 void clearOperationActions() {
973 /// findRepresentativeClass - Return the largest legal super-reg register class
974 /// of the register class for the specified type and its associated "cost".
975 virtual std::pair<const TargetRegisterClass*, uint8_t>
976 findRepresentativeClass(MVT VT) const;
978 /// computeRegisterProperties - Once all of the register classes are added,
979 /// this allows us to compute derived properties we expose.
980 void computeRegisterProperties();
982 /// setOperationAction - Indicate that the specified operation does not work
983 /// with the specified type and indicate what to do about it.
984 void setOperationAction(unsigned Op, MVT VT,
985 LegalizeAction Action) {
986 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!");
987 OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action;
990 /// setLoadExtAction - Indicate that the specified load with extension does
991 /// not work with the specified type and indicate what to do about it.
992 void setLoadExtAction(unsigned ExtType, MVT VT,
993 LegalizeAction Action) {
994 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE &&
995 "Table isn't big enough!");
996 LoadExtActions[VT.SimpleTy][ExtType] = (uint8_t)Action;
999 /// setTruncStoreAction - Indicate that the specified truncating store does
1000 /// not work with the specified type and indicate what to do about it.
1001 void setTruncStoreAction(MVT ValVT, MVT MemVT,
1002 LegalizeAction Action) {
1003 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE &&
1004 "Table isn't big enough!");
1005 TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action;
1008 /// setIndexedLoadAction - Indicate that the specified indexed load does or
1009 /// does not work with the specified type and indicate what to do abort
1010 /// it. NOTE: All indexed mode loads are initialized to Expand in
1011 /// TargetLowering.cpp
1012 void setIndexedLoadAction(unsigned IdxMode, MVT VT,
1013 LegalizeAction Action) {
1014 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE &&
1015 (unsigned)Action < 0xf && "Table isn't big enough!");
1016 // Load action are kept in the upper half.
1017 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0;
1018 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4;
1021 /// setIndexedStoreAction - Indicate that the specified indexed store does or
1022 /// does not work with the specified type and indicate what to do about
1023 /// it. NOTE: All indexed mode stores are initialized to Expand in
1024 /// TargetLowering.cpp
1025 void setIndexedStoreAction(unsigned IdxMode, MVT VT,
1026 LegalizeAction Action) {
1027 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE &&
1028 (unsigned)Action < 0xf && "Table isn't big enough!");
1029 // Store action are kept in the lower half.
1030 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f;
1031 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action);
1034 /// setCondCodeAction - Indicate that the specified condition code is or isn't
1035 /// supported on the target and indicate what to do about it.
1036 void setCondCodeAction(ISD::CondCode CC, MVT VT,
1037 LegalizeAction Action) {
1038 assert(VT < MVT::LAST_VALUETYPE &&
1039 (unsigned)CC < array_lengthof(CondCodeActions) &&
1040 "Table isn't big enough!");
1041 /// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 64bit
1042 /// value and the upper 27 bits index into the second dimension of the
1043 /// array to select what 64bit value to use.
1044 CondCodeActions[(unsigned)CC][VT.SimpleTy >> 5]
1045 &= ~(uint64_t(3UL) << (VT.SimpleTy & 0x1F)*2);
1046 CondCodeActions[(unsigned)CC][VT.SimpleTy >> 5]
1047 |= (uint64_t)Action << (VT.SimpleTy & 0x1F)*2;
1050 /// AddPromotedToType - If Opc/OrigVT is specified as being promoted, the
1051 /// promotion code defaults to trying a larger integer/fp until it can find
1052 /// one that works. If that default is insufficient, this method can be used
1053 /// by the target to override the default.
1054 void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) {
1055 PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy;
1058 /// setTargetDAGCombine - Targets should invoke this method for each target
1059 /// independent node that they want to provide a custom DAG combiner for by
1060 /// implementing the PerformDAGCombine virtual method.
1061 void setTargetDAGCombine(ISD::NodeType NT) {
1062 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
1063 TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7);
1066 /// setJumpBufSize - Set the target's required jmp_buf buffer size (in
1067 /// bytes); default is 200
1068 void setJumpBufSize(unsigned Size) {
1072 /// setJumpBufAlignment - Set the target's required jmp_buf buffer
1073 /// alignment (in bytes); default is 0
1074 void setJumpBufAlignment(unsigned Align) {
1075 JumpBufAlignment = Align;
1078 /// setMinFunctionAlignment - Set the target's minimum function alignment (in
1080 void setMinFunctionAlignment(unsigned Align) {
1081 MinFunctionAlignment = Align;
1084 /// setPrefFunctionAlignment - Set the target's preferred function alignment.
1085 /// This should be set if there is a performance benefit to
1086 /// higher-than-minimum alignment (in log2(bytes))
1087 void setPrefFunctionAlignment(unsigned Align) {
1088 PrefFunctionAlignment = Align;
1091 /// setPrefLoopAlignment - Set the target's preferred loop alignment. Default
1092 /// alignment is zero, it means the target does not care about loop alignment.
1093 /// The alignment is specified in log2(bytes).
1094 void setPrefLoopAlignment(unsigned Align) {
1095 PrefLoopAlignment = Align;
1098 /// setMinStackArgumentAlignment - Set the minimum stack alignment of an
1099 /// argument (in log2(bytes)).
1100 void setMinStackArgumentAlignment(unsigned Align) {
1101 MinStackArgumentAlignment = Align;
1104 /// setShouldFoldAtomicFences - Set if the target's implementation of the
1105 /// atomic operation intrinsics includes locking. Default is false.
1106 void setShouldFoldAtomicFences(bool fold) {
1107 ShouldFoldAtomicFences = fold;
1110 /// setInsertFencesForAtomic - Set if the DAG builder should
1111 /// automatically insert fences and reduce the order of atomic memory
1112 /// operations to Monotonic.
1113 void setInsertFencesForAtomic(bool fence) {
1114 InsertFencesForAtomic = fence;
1118 //===--------------------------------------------------------------------===//
1119 // Addressing mode description hooks (used by LSR etc).
1122 /// GetAddrModeArguments - CodeGenPrepare sinks address calculations into the
1123 /// same BB as Load/Store instructions reading the address. This allows as
1124 /// much computation as possible to be done in the address mode for that
1125 /// operand. This hook lets targets also pass back when this should be done
1126 /// on intrinsics which load/store.
1127 virtual bool GetAddrModeArguments(IntrinsicInst *I,
1128 SmallVectorImpl<Value*> &Ops,
1129 Type *&AccessTy) const {
1133 /// AddrMode - This represents an addressing mode of:
1134 /// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
1135 /// If BaseGV is null, there is no BaseGV.
1136 /// If BaseOffs is zero, there is no base offset.
1137 /// If HasBaseReg is false, there is no base register.
1138 /// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with
1142 GlobalValue *BaseGV;
1146 AddrMode() : BaseGV(0), BaseOffs(0), HasBaseReg(false), Scale(0) {}
1149 /// isLegalAddressingMode - Return true if the addressing mode represented by
1150 /// AM is legal for this target, for a load/store of the specified type.
1151 /// The type may be VoidTy, in which case only return true if the addressing
1152 /// mode is legal for a load/store of any legal type.
1153 /// TODO: Handle pre/postinc as well.
1154 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const;
1156 /// isLegalICmpImmediate - Return true if the specified immediate is legal
1157 /// icmp immediate, that is the target has icmp instructions which can compare
1158 /// a register against the immediate without having to materialize the
1159 /// immediate into a register.
1160 virtual bool isLegalICmpImmediate(int64_t) const {
1164 /// isLegalAddImmediate - Return true if the specified immediate is legal
1165 /// add immediate, that is the target has add instructions which can add
1166 /// a register with the immediate without having to materialize the
1167 /// immediate into a register.
1168 virtual bool isLegalAddImmediate(int64_t) const {
1172 /// isTruncateFree - Return true if it's free to truncate a value of
1173 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
1174 /// register EAX to i16 by referencing its sub-register AX.
1175 virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const {
1179 virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const {
1183 /// isZExtFree - Return true if any actual instruction that defines a
1184 /// value of type Ty1 implicitly zero-extends the value to Ty2 in the result
1185 /// register. This does not necessarily include registers defined in
1186 /// unknown ways, such as incoming arguments, or copies from unknown
1187 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
1188 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
1189 /// all instructions that define 32-bit values implicit zero-extend the
1190 /// result out to 64 bits.
1191 virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const {
1195 virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const {
1199 /// isZExtFree - Return true if zero-extending the specific node Val to type
1200 /// VT2 is free (either because it's implicitly zero-extended such as ARM
1201 /// ldrb / ldrh or because it's folded such as X86 zero-extending loads).
1202 virtual bool isZExtFree(SDValue Val, EVT VT2) const {
1203 return isZExtFree(Val.getValueType(), VT2);
1206 /// isFNegFree - Return true if an fneg operation is free to the point where
1207 /// it is never worthwhile to replace it with a bitwise operation.
1208 virtual bool isFNegFree(EVT) const {
1212 /// isFAbsFree - Return true if an fneg operation is free to the point where
1213 /// it is never worthwhile to replace it with a bitwise operation.
1214 virtual bool isFAbsFree(EVT) const {
1218 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
1219 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
1220 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
1221 /// is expanded to mul + add.
1222 virtual bool isFMAFasterThanMulAndAdd(EVT) const {
1226 /// isNarrowingProfitable - Return true if it's profitable to narrow
1227 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
1228 /// from i32 to i8 but not from i32 to i16.
1229 virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const {
1233 //===--------------------------------------------------------------------===//
1234 // Runtime Library hooks
1237 /// setLibcallName - Rename the default libcall routine name for the specified
1239 void setLibcallName(RTLIB::Libcall Call, const char *Name) {
1240 LibcallRoutineNames[Call] = Name;
1243 /// getLibcallName - Get the libcall routine name for the specified libcall.
1245 const char *getLibcallName(RTLIB::Libcall Call) const {
1246 return LibcallRoutineNames[Call];
1249 /// setCmpLibcallCC - Override the default CondCode to be used to test the
1250 /// result of the comparison libcall against zero.
1251 void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) {
1252 CmpLibcallCCs[Call] = CC;
1255 /// getCmpLibcallCC - Get the CondCode that's to be used to test the result of
1256 /// the comparison libcall against zero.
1257 ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const {
1258 return CmpLibcallCCs[Call];
1261 /// setLibcallCallingConv - Set the CallingConv that should be used for the
1262 /// specified libcall.
1263 void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) {
1264 LibcallCallingConvs[Call] = CC;
1267 /// getLibcallCallingConv - Get the CallingConv that should be used for the
1268 /// specified libcall.
1269 CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const {
1270 return LibcallCallingConvs[Call];
1274 const TargetMachine &TM;
1275 const DataLayout *TD;
1276 const TargetLoweringObjectFile &TLOF;
1278 /// PointerTy - The type to use for pointers for the default address space,
1279 /// usually i32 or i64.
1283 /// IsLittleEndian - True if this is a little endian target.
1285 bool IsLittleEndian;
1287 /// SelectIsExpensive - Tells the code generator not to expand operations
1288 /// into sequences that use the select operations if possible.
1289 bool SelectIsExpensive;
1291 /// IntDivIsCheap - Tells the code generator not to expand integer divides by
1292 /// constants into a sequence of muls, adds, and shifts. This is a hack until
1293 /// a real cost model is in place. If we ever optimize for size, this will be
1294 /// set to true unconditionally.
1297 /// BypassSlowDivMap - Tells the code generator to bypass slow divide or
1298 /// remainder instructions. For example, BypassSlowDivWidths[32,8] tells the
1299 /// code generator to bypass 32-bit integer div/rem with an 8-bit unsigned
1300 /// integer div/rem when the operands are positive and less than 256.
1301 DenseMap <unsigned int, unsigned int> BypassSlowDivWidths;
1303 /// Pow2DivIsCheap - Tells the code generator that it shouldn't generate
1304 /// srl/add/sra for a signed divide by power of two, and let the target handle
1306 bool Pow2DivIsCheap;
1308 /// JumpIsExpensive - Tells the code generator that it shouldn't generate
1309 /// extra flow control instructions and should attempt to combine flow
1310 /// control instructions via predication.
1311 bool JumpIsExpensive;
1313 /// UseUnderscoreSetJmp - This target prefers to use _setjmp to implement
1314 /// llvm.setjmp. Defaults to false.
1315 bool UseUnderscoreSetJmp;
1317 /// UseUnderscoreLongJmp - This target prefers to use _longjmp to implement
1318 /// llvm.longjmp. Defaults to false.
1319 bool UseUnderscoreLongJmp;
1321 /// SupportJumpTables - Whether the target can generate code for jumptables.
1322 /// If it's not true, then each jumptable must be lowered into if-then-else's.
1323 bool SupportJumpTables;
1325 /// MinimumJumpTableEntries - Number of blocks threshold to use jump tables.
1326 int MinimumJumpTableEntries;
1328 /// BooleanContents - Information about the contents of the high-bits in
1329 /// boolean values held in a type wider than i1. See getBooleanContents.
1330 BooleanContent BooleanContents;
1331 /// BooleanVectorContents - Information about the contents of the high-bits
1332 /// in boolean vector values when the element type is wider than i1. See
1333 /// getBooleanContents.
1334 BooleanContent BooleanVectorContents;
1336 /// SchedPreferenceInfo - The target scheduling preference: shortest possible
1337 /// total cycles or lowest register usage.
1338 Sched::Preference SchedPreferenceInfo;
1340 /// JumpBufSize - The size, in bytes, of the target's jmp_buf buffers
1341 unsigned JumpBufSize;
1343 /// JumpBufAlignment - The alignment, in bytes, of the target's jmp_buf
1345 unsigned JumpBufAlignment;
1347 /// MinStackArgumentAlignment - The minimum alignment that any argument
1348 /// on the stack needs to have.
1350 unsigned MinStackArgumentAlignment;
1352 /// MinFunctionAlignment - The minimum function alignment (used when
1353 /// optimizing for size, and to prevent explicitly provided alignment
1354 /// from leading to incorrect code).
1356 unsigned MinFunctionAlignment;
1358 /// PrefFunctionAlignment - The preferred function alignment (used when
1359 /// alignment unspecified and optimizing for speed).
1361 unsigned PrefFunctionAlignment;
1363 /// PrefLoopAlignment - The preferred loop alignment.
1365 unsigned PrefLoopAlignment;
1367 /// ShouldFoldAtomicFences - Whether fencing MEMBARRIER instructions should
1368 /// be folded into the enclosed atomic intrinsic instruction by the
1370 bool ShouldFoldAtomicFences;
1372 /// InsertFencesForAtomic - Whether the DAG builder should automatically
1373 /// insert fences and reduce ordering for atomics. (This will be set for
1374 /// for most architectures with weak memory ordering.)
1375 bool InsertFencesForAtomic;
1377 /// StackPointerRegisterToSaveRestore - If set to a physical register, this
1378 /// specifies the register that llvm.savestack/llvm.restorestack should save
1380 unsigned StackPointerRegisterToSaveRestore;
1382 /// ExceptionPointerRegister - If set to a physical register, this specifies
1383 /// the register that receives the exception address on entry to a landing
1385 unsigned ExceptionPointerRegister;
1387 /// ExceptionSelectorRegister - If set to a physical register, this specifies
1388 /// the register that receives the exception typeid on entry to a landing
1390 unsigned ExceptionSelectorRegister;
1392 /// RegClassForVT - This indicates the default register class to use for
1393 /// each ValueType the target supports natively.
1394 const TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE];
1395 unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE];
1396 MVT RegisterTypeForVT[MVT::LAST_VALUETYPE];
1398 /// RepRegClassForVT - This indicates the "representative" register class to
1399 /// use for each ValueType the target supports natively. This information is
1400 /// used by the scheduler to track register pressure. By default, the
1401 /// representative register class is the largest legal super-reg register
1402 /// class of the register class of the specified type. e.g. On x86, i8, i16,
1403 /// and i32's representative class would be GR32.
1404 const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE];
1406 /// RepRegClassCostForVT - This indicates the "cost" of the "representative"
1407 /// register class for each ValueType. The cost is used by the scheduler to
1408 /// approximate register pressure.
1409 uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE];
1411 /// TransformToType - For any value types we are promoting or expanding, this
1412 /// contains the value type that we are changing to. For Expanded types, this
1413 /// contains one step of the expand (e.g. i64 -> i32), even if there are
1414 /// multiple steps required (e.g. i64 -> i16). For types natively supported
1415 /// by the system, this holds the same type (e.g. i32 -> i32).
1416 MVT TransformToType[MVT::LAST_VALUETYPE];
1418 /// OpActions - For each operation and each value type, keep a LegalizeAction
1419 /// that indicates how instruction selection should deal with the operation.
1420 /// Most operations are Legal (aka, supported natively by the target), but
1421 /// operations that are not should be described. Note that operations on
1422 /// non-legal value types are not described here.
1423 uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END];
1425 /// LoadExtActions - For each load extension type and each value type,
1426 /// keep a LegalizeAction that indicates how instruction selection should deal
1427 /// with a load of a specific value type and extension type.
1428 uint8_t LoadExtActions[MVT::LAST_VALUETYPE][ISD::LAST_LOADEXT_TYPE];
1430 /// TruncStoreActions - For each value type pair keep a LegalizeAction that
1431 /// indicates whether a truncating store of a specific value type and
1432 /// truncating type is legal.
1433 uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE];
1435 /// IndexedModeActions - For each indexed mode and each value type,
1436 /// keep a pair of LegalizeAction that indicates how instruction
1437 /// selection should deal with the load / store. The first dimension is the
1438 /// value_type for the reference. The second dimension represents the various
1439 /// modes for load store.
1440 uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE];
1442 /// CondCodeActions - For each condition code (ISD::CondCode) keep a
1443 /// LegalizeAction that indicates how instruction selection should
1444 /// deal with the condition code.
1445 /// Because each CC action takes up 2 bits, we need to have the array size
1446 /// be large enough to fit all of the value types. This can be done by
1447 /// dividing the MVT::LAST_VALUETYPE by 32 and adding one.
1448 uint64_t CondCodeActions[ISD::SETCC_INVALID][(MVT::LAST_VALUETYPE / 32) + 1];
1450 ValueTypeActionImpl ValueTypeActions;
1454 getTypeConversion(LLVMContext &Context, EVT VT) const {
1455 // If this is a simple type, use the ComputeRegisterProp mechanism.
1456 if (VT.isSimple()) {
1457 MVT SVT = VT.getSimpleVT();
1458 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
1459 MVT NVT = TransformToType[SVT.SimpleTy];
1460 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
1464 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)
1465 && "Promote may not follow Expand or Promote");
1467 if (LA == TypeSplitVector)
1468 return LegalizeKind(LA, EVT::getVectorVT(Context,
1469 SVT.getVectorElementType(),
1470 SVT.getVectorNumElements()/2));
1471 if (LA == TypeScalarizeVector)
1472 return LegalizeKind(LA, SVT.getVectorElementType());
1473 return LegalizeKind(LA, NVT);
1476 // Handle Extended Scalar Types.
1477 if (!VT.isVector()) {
1478 assert(VT.isInteger() && "Float types must be simple");
1479 unsigned BitSize = VT.getSizeInBits();
1480 // First promote to a power-of-two size, then expand if necessary.
1481 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
1482 EVT NVT = VT.getRoundIntegerType(Context);
1483 assert(NVT != VT && "Unable to round integer VT");
1484 LegalizeKind NextStep = getTypeConversion(Context, NVT);
1485 // Avoid multi-step promotion.
1486 if (NextStep.first == TypePromoteInteger) return NextStep;
1487 // Return rounded integer type.
1488 return LegalizeKind(TypePromoteInteger, NVT);
1491 return LegalizeKind(TypeExpandInteger,
1492 EVT::getIntegerVT(Context, VT.getSizeInBits()/2));
1495 // Handle vector types.
1496 unsigned NumElts = VT.getVectorNumElements();
1497 EVT EltVT = VT.getVectorElementType();
1499 // Vectors with only one element are always scalarized.
1501 return LegalizeKind(TypeScalarizeVector, EltVT);
1503 // Try to widen vector elements until a legal type is found.
1504 if (EltVT.isInteger()) {
1505 // Vectors with a number of elements that is not a power of two are always
1506 // widened, for example <3 x float> -> <4 x float>.
1507 if (!VT.isPow2VectorType()) {
1508 NumElts = (unsigned)NextPowerOf2(NumElts);
1509 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
1510 return LegalizeKind(TypeWidenVector, NVT);
1513 // Examine the element type.
1514 LegalizeKind LK = getTypeConversion(Context, EltVT);
1516 // If type is to be expanded, split the vector.
1517 // <4 x i140> -> <2 x i140>
1518 if (LK.first == TypeExpandInteger)
1519 return LegalizeKind(TypeSplitVector,
1520 EVT::getVectorVT(Context, EltVT, NumElts / 2));
1522 // Promote the integer element types until a legal vector type is found
1523 // or until the element integer type is too big. If a legal type was not
1524 // found, fallback to the usual mechanism of widening/splitting the
1526 EVT OldEltVT = EltVT;
1528 // Increase the bitwidth of the element to the next pow-of-two
1529 // (which is greater than 8 bits).
1530 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()
1531 ).getRoundIntegerType(Context);
1533 // Stop trying when getting a non-simple element type.
1534 // Note that vector elements may be greater than legal vector element
1535 // types. Example: X86 XMM registers hold 64bit element on 32bit systems.
1536 if (!EltVT.isSimple()) break;
1538 // Build a new vector type and check if it is legal.
1539 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1540 // Found a legal promoted vector type.
1541 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
1542 return LegalizeKind(TypePromoteInteger,
1543 EVT::getVectorVT(Context, EltVT, NumElts));
1546 // Reset the type to the unexpanded type if we did not find a legal vector
1547 // type with a promoted vector element type.
1551 // Try to widen the vector until a legal type is found.
1552 // If there is no wider legal type, split the vector.
1554 // Round up to the next power of 2.
1555 NumElts = (unsigned)NextPowerOf2(NumElts);
1557 // If there is no simple vector type with this many elements then there
1558 // cannot be a larger legal vector type. Note that this assumes that
1559 // there are no skipped intermediate vector types in the simple types.
1560 if (!EltVT.isSimple()) break;
1561 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1562 if (LargerVector == MVT()) break;
1564 // If this type is legal then widen the vector.
1565 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
1566 return LegalizeKind(TypeWidenVector, LargerVector);
1569 // Widen odd vectors to next power of two.
1570 if (!VT.isPow2VectorType()) {
1571 EVT NVT = VT.getPow2VectorType(Context);
1572 return LegalizeKind(TypeWidenVector, NVT);
1575 // Vectors with illegal element types are expanded.
1576 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
1577 return LegalizeKind(TypeSplitVector, NVT);
1581 std::vector<std::pair<MVT, const TargetRegisterClass*> > AvailableRegClasses;
1583 /// TargetDAGCombineArray - Targets can specify ISD nodes that they would
1584 /// like PerformDAGCombine callbacks for by calling setTargetDAGCombine(),
1585 /// which sets a bit in this array.
1587 TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT];
1589 /// PromoteToType - For operations that must be promoted to a specific type,
1590 /// this holds the destination type. This map should be sparse, so don't hold
1593 /// Targets add entries to this map with AddPromotedToType(..), clients access
1594 /// this with getTypeToPromoteTo(..).
1595 std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType>
1598 /// LibcallRoutineNames - Stores the name each libcall.
1600 const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL];
1602 /// CmpLibcallCCs - The ISD::CondCode that should be used to test the result
1603 /// of each of the comparison libcall against zero.
1604 ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL];
1606 /// LibcallCallingConvs - Stores the CallingConv that should be used for each
1608 CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL];
1611 /// When lowering \@llvm.memset this field specifies the maximum number of
1612 /// store operations that may be substituted for the call to memset. Targets
1613 /// must set this value based on the cost threshold for that target. Targets
1614 /// should assume that the memset will be done using as many of the largest
1615 /// store operations first, followed by smaller ones, if necessary, per
1616 /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine
1617 /// with 16-bit alignment would result in four 2-byte stores and one 1-byte
1618 /// store. This only applies to setting a constant array of a constant size.
1619 /// @brief Specify maximum number of store instructions per memset call.
1620 unsigned MaxStoresPerMemset;
1622 /// Maximum number of stores operations that may be substituted for the call
1623 /// to memset, used for functions with OptSize attribute.
1624 unsigned MaxStoresPerMemsetOptSize;
1626 /// When lowering \@llvm.memcpy this field specifies the maximum number of
1627 /// store operations that may be substituted for a call to memcpy. Targets
1628 /// must set this value based on the cost threshold for that target. Targets
1629 /// should assume that the memcpy will be done using as many of the largest
1630 /// store operations first, followed by smaller ones, if necessary, per
1631 /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine
1632 /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store
1633 /// and one 1-byte store. This only applies to copying a constant array of
1635 /// @brief Specify maximum bytes of store instructions per memcpy call.
1636 unsigned MaxStoresPerMemcpy;
1638 /// Maximum number of store operations that may be substituted for a call
1639 /// to memcpy, used for functions with OptSize attribute.
1640 unsigned MaxStoresPerMemcpyOptSize;
1642 /// When lowering \@llvm.memmove this field specifies the maximum number of
1643 /// store instructions that may be substituted for a call to memmove. Targets
1644 /// must set this value based on the cost threshold for that target. Targets
1645 /// should assume that the memmove will be done using as many of the largest
1646 /// store operations first, followed by smaller ones, if necessary, per
1647 /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine
1648 /// with 8-bit alignment would result in nine 1-byte stores. This only
1649 /// applies to copying a constant array of constant size.
1650 /// @brief Specify maximum bytes of store instructions per memmove call.
1651 unsigned MaxStoresPerMemmove;
1653 /// Maximum number of store instructions that may be substituted for a call
1654 /// to memmove, used for functions with OpSize attribute.
1655 unsigned MaxStoresPerMemmoveOptSize;
1657 /// PredictableSelectIsExpensive - Tells the code generator that select is
1658 /// more expensive than a branch if the branch is usually predicted right.
1659 bool PredictableSelectIsExpensive;
1662 /// isLegalRC - Return true if the value types that can be represented by the
1663 /// specified register class are all legal.
1664 bool isLegalRC(const TargetRegisterClass *RC) const;
1667 //===----------------------------------------------------------------------===//
1668 /// TargetLowering - This class defines information used to lower LLVM code to
1669 /// legal SelectionDAG operators that the target instruction selector can accept
1672 /// This class also defines callbacks that targets must implement to lower
1673 /// target-specific constructs to SelectionDAG operators.
1675 class TargetLowering : public TargetLoweringBase {
1676 TargetLowering(const TargetLowering&) LLVM_DELETED_FUNCTION;
1677 void operator=(const TargetLowering&) LLVM_DELETED_FUNCTION;
1680 /// NOTE: The constructor takes ownership of TLOF.
1681 explicit TargetLowering(const TargetMachine &TM,
1682 const TargetLoweringObjectFile *TLOF);
1684 /// getPreIndexedAddressParts - returns true by value, base pointer and
1685 /// offset pointer and addressing mode by reference if the node's address
1686 /// can be legally represented as pre-indexed load / store address.
1687 virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/,
1688 SDValue &/*Offset*/,
1689 ISD::MemIndexedMode &/*AM*/,
1690 SelectionDAG &/*DAG*/) const {
1694 /// getPostIndexedAddressParts - returns true by value, base pointer and
1695 /// offset pointer and addressing mode by reference if this node can be
1696 /// combined with a load / store to form a post-indexed load / store.
1697 virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/,
1698 SDValue &/*Base*/, SDValue &/*Offset*/,
1699 ISD::MemIndexedMode &/*AM*/,
1700 SelectionDAG &/*DAG*/) const {
1704 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1705 /// current function. The returned value is a member of the
1706 /// MachineJumpTableInfo::JTEntryKind enum.
1707 virtual unsigned getJumpTableEncoding() const;
1709 virtual const MCExpr *
1710 LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/,
1711 const MachineBasicBlock * /*MBB*/, unsigned /*uid*/,
1712 MCContext &/*Ctx*/) const {
1713 llvm_unreachable("Need to implement this hook if target has custom JTIs");
1716 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1718 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
1719 SelectionDAG &DAG) const;
1721 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1722 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1724 virtual const MCExpr *
1725 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1726 unsigned JTI, MCContext &Ctx) const;
1728 /// isOffsetFoldingLegal - Return true if folding a constant offset
1729 /// with the given GlobalAddress is legal. It is frequently not legal in
1730 /// PIC relocation models.
1731 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
1733 bool isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
1734 SDValue &Chain) const;
1736 void softenSetCCOperands(SelectionDAG &DAG, EVT VT,
1737 SDValue &NewLHS, SDValue &NewRHS,
1738 ISD::CondCode &CCCode, DebugLoc DL) const;
1740 SDValue makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC, EVT RetVT,
1741 const SDValue *Ops, unsigned NumOps,
1742 bool isSigned, DebugLoc dl) const;
1744 //===--------------------------------------------------------------------===//
1745 // TargetLowering Optimization Methods
1748 /// TargetLoweringOpt - A convenience struct that encapsulates a DAG, and two
1749 /// SDValues for returning information from TargetLowering to its clients
1750 /// that want to combine
1751 struct TargetLoweringOpt {
1758 explicit TargetLoweringOpt(SelectionDAG &InDAG,
1760 DAG(InDAG), LegalTys(LT), LegalOps(LO) {}
1762 bool LegalTypes() const { return LegalTys; }
1763 bool LegalOperations() const { return LegalOps; }
1765 bool CombineTo(SDValue O, SDValue N) {
1771 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1772 /// specified instruction is a constant integer. If so, check to see if
1773 /// there are any bits set in the constant that are not demanded. If so,
1774 /// shrink the constant and return true.
1775 bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded);
1777 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
1778 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
1779 /// cast, but it could be generalized for targets with other types of
1780 /// implicit widening casts.
1781 bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded,
1785 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
1786 /// DemandedMask bits of the result of Op are ever used downstream. If we can
1787 /// use this information to simplify Op, create a new simplified DAG node and
1788 /// return true, returning the original and new nodes in Old and New.
1789 /// Otherwise, analyze the expression and return a mask of KnownOne and
1790 /// KnownZero bits for the expression (used to simplify the caller).
1791 /// The KnownZero/One bits may only be accurate for those bits in the
1793 bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask,
1794 APInt &KnownZero, APInt &KnownOne,
1795 TargetLoweringOpt &TLO, unsigned Depth = 0) const;
1797 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
1798 /// Mask are known to be either zero or one and return them in the
1799 /// KnownZero/KnownOne bitsets.
1800 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
1803 const SelectionDAG &DAG,
1804 unsigned Depth = 0) const;
1806 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1807 /// targets that want to expose additional information about sign bits to the
1809 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
1810 unsigned Depth = 0) const;
1812 struct DAGCombinerInfo {
1813 void *DC; // The DAG Combiner object.
1815 bool CalledByLegalizer;
1819 DAGCombinerInfo(SelectionDAG &dag, CombineLevel level, bool cl, void *dc)
1820 : DC(dc), Level(level), CalledByLegalizer(cl), DAG(dag) {}
1822 bool isBeforeLegalize() const { return Level == BeforeLegalizeTypes; }
1823 bool isBeforeLegalizeOps() const { return Level < AfterLegalizeVectorOps; }
1824 bool isAfterLegalizeVectorOps() const {
1825 return Level == AfterLegalizeDAG;
1827 CombineLevel getDAGCombineLevel() { return Level; }
1828 bool isCalledByLegalizer() const { return CalledByLegalizer; }
1830 void AddToWorklist(SDNode *N);
1831 void RemoveFromWorklist(SDNode *N);
1832 SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To,
1834 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true);
1835 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true);
1837 void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO);
1840 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1841 /// and cc. If it is unable to simplify it, return a null SDValue.
1842 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1843 ISD::CondCode Cond, bool foldBooleans,
1844 DAGCombinerInfo &DCI, DebugLoc dl) const;
1846 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1847 /// node is a GlobalAddress + offset.
1849 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
1851 /// PerformDAGCombine - This method will be invoked for all target nodes and
1852 /// for any target-independent nodes that the target has registered with
1855 /// The semantics are as follows:
1857 /// SDValue.Val == 0 - No change was made
1858 /// SDValue.Val == N - N was replaced, is dead, and is already handled.
1859 /// otherwise - N should be replaced by the returned Operand.
1861 /// In addition, methods provided by DAGCombinerInfo may be used to perform
1862 /// more complex transformations.
1864 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
1866 /// isTypeDesirableForOp - Return true if the target has native support for
1867 /// the specified value type and it is 'desirable' to use the type for the
1868 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
1869 /// instruction encodings are longer and some i16 instructions are slow.
1870 virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const {
1871 // By default, assume all legal types are desirable.
1872 return isTypeLegal(VT);
1875 /// isDesirableToPromoteOp - Return true if it is profitable for dag combiner
1876 /// to transform a floating point op of specified opcode to a equivalent op of
1877 /// an integer type. e.g. f32 load -> i32 load can be profitable on ARM.
1878 virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/,
1883 /// IsDesirableToPromoteOp - This method query the target whether it is
1884 /// beneficial for dag combiner to promote the specified node. If true, it
1885 /// should return the desired promotion type by reference.
1886 virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const {
1890 //===--------------------------------------------------------------------===//
1891 // Lowering methods - These methods must be implemented by targets so that
1892 // the SelectionDAGBuilder code knows how to lower these.
1895 /// LowerFormalArguments - This hook must be implemented to lower the
1896 /// incoming (formal) arguments, described by the Ins array, into the
1897 /// specified DAG. The implementation should fill in the InVals array
1898 /// with legal-type argument values, and return the resulting token
1902 LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/,
1904 const SmallVectorImpl<ISD::InputArg> &/*Ins*/,
1905 DebugLoc /*dl*/, SelectionDAG &/*DAG*/,
1906 SmallVectorImpl<SDValue> &/*InVals*/) const {
1907 llvm_unreachable("Not Implemented");
1910 struct ArgListEntry {
1919 bool isReturned : 1;
1922 ArgListEntry() : isSExt(false), isZExt(false), isInReg(false),
1923 isSRet(false), isNest(false), isByVal(false), isReturned(false),
1926 typedef std::vector<ArgListEntry> ArgListTy;
1928 /// CallLoweringInfo - This structure contains all information that is
1929 /// necessary for lowering calls. It is passed to TLI::LowerCallTo when the
1930 /// SelectionDAG builder needs to lower a call, and targets will see this
1931 /// struct in their LowerCall implementation.
1932 struct CallLoweringInfo {
1939 bool DoesNotReturn : 1;
1940 bool IsReturnValueUsed : 1;
1942 // IsTailCall should be modified by implementations of
1943 // TargetLowering::LowerCall that perform tail call conversions.
1946 unsigned NumFixedArgs;
1947 CallingConv::ID CallConv;
1952 ImmutableCallSite *CS;
1953 SmallVector<ISD::OutputArg, 32> Outs;
1954 SmallVector<SDValue, 32> OutVals;
1955 SmallVector<ISD::InputArg, 32> Ins;
1958 /// CallLoweringInfo - Constructs a call lowering context based on the
1959 /// ImmutableCallSite \p cs.
1960 CallLoweringInfo(SDValue chain, Type *retTy,
1961 FunctionType *FTy, bool isTailCall, SDValue callee,
1962 ArgListTy &args, SelectionDAG &dag, DebugLoc dl,
1963 ImmutableCallSite &cs)
1964 : Chain(chain), RetTy(retTy), RetSExt(cs.paramHasAttr(0, Attribute::SExt)),
1965 RetZExt(cs.paramHasAttr(0, Attribute::ZExt)), IsVarArg(FTy->isVarArg()),
1966 IsInReg(cs.paramHasAttr(0, Attribute::InReg)),
1967 DoesNotReturn(cs.doesNotReturn()),
1968 IsReturnValueUsed(!cs.getInstruction()->use_empty()),
1969 IsTailCall(isTailCall), NumFixedArgs(FTy->getNumParams()),
1970 CallConv(cs.getCallingConv()), Callee(callee), Args(args), DAG(dag),
1973 /// CallLoweringInfo - Constructs a call lowering context based on the
1974 /// provided call information.
1975 CallLoweringInfo(SDValue chain, Type *retTy, bool retSExt, bool retZExt,
1976 bool isVarArg, bool isInReg, unsigned numFixedArgs,
1977 CallingConv::ID callConv, bool isTailCall,
1978 bool doesNotReturn, bool isReturnValueUsed, SDValue callee,
1979 ArgListTy &args, SelectionDAG &dag, DebugLoc dl)
1980 : Chain(chain), RetTy(retTy), RetSExt(retSExt), RetZExt(retZExt),
1981 IsVarArg(isVarArg), IsInReg(isInReg), DoesNotReturn(doesNotReturn),
1982 IsReturnValueUsed(isReturnValueUsed), IsTailCall(isTailCall),
1983 NumFixedArgs(numFixedArgs), CallConv(callConv), Callee(callee),
1984 Args(args), DAG(dag), DL(dl), CS(NULL) {}
1987 /// LowerCallTo - This function lowers an abstract call to a function into an
1988 /// actual call. This returns a pair of operands. The first element is the
1989 /// return value for the function (if RetTy is not VoidTy). The second
1990 /// element is the outgoing token chain. It calls LowerCall to do the actual
1992 std::pair<SDValue, SDValue> LowerCallTo(CallLoweringInfo &CLI) const;
1994 /// LowerCall - This hook must be implemented to lower calls into the
1995 /// the specified DAG. The outgoing arguments to the call are described
1996 /// by the Outs array, and the values to be returned by the call are
1997 /// described by the Ins array. The implementation should fill in the
1998 /// InVals array with legal-type return values from the call, and return
1999 /// the resulting token chain value.
2001 LowerCall(CallLoweringInfo &/*CLI*/,
2002 SmallVectorImpl<SDValue> &/*InVals*/) const {
2003 llvm_unreachable("Not Implemented");
2006 /// HandleByVal - Target-specific cleanup for formal ByVal parameters.
2007 virtual void HandleByVal(CCState *, unsigned &, unsigned) const {}
2009 /// CanLowerReturn - This hook should be implemented to check whether the
2010 /// return values described by the Outs array can fit into the return
2011 /// registers. If false is returned, an sret-demotion is performed.
2013 virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/,
2014 MachineFunction &/*MF*/, bool /*isVarArg*/,
2015 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/,
2016 LLVMContext &/*Context*/) const
2018 // Return true by default to get preexisting behavior.
2022 /// LowerReturn - This hook must be implemented to lower outgoing
2023 /// return values, described by the Outs array, into the specified
2024 /// DAG. The implementation should return the resulting token chain
2028 LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/,
2030 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/,
2031 const SmallVectorImpl<SDValue> &/*OutVals*/,
2032 DebugLoc /*dl*/, SelectionDAG &/*DAG*/) const {
2033 llvm_unreachable("Not Implemented");
2036 /// isUsedByReturnOnly - Return true if result of the specified node is used
2037 /// by a return node only. It also compute and return the input chain for the
2039 /// This is used to determine whether it is possible
2040 /// to codegen a libcall as tail call at legalization time.
2041 virtual bool isUsedByReturnOnly(SDNode *, SDValue &Chain) const {
2045 /// mayBeEmittedAsTailCall - Return true if the target may be able emit the
2046 /// call instruction as a tail call. This is used by optimization passes to
2047 /// determine if it's profitable to duplicate return instructions to enable
2048 /// tailcall optimization.
2049 virtual bool mayBeEmittedAsTailCall(CallInst *) const {
2053 /// getTypeForExtArgOrReturn - Return the type that should be used to zero or
2054 /// sign extend a zeroext/signext integer argument or return value.
2055 /// FIXME: Most C calling convention requires the return type to be promoted,
2056 /// but this is not true all the time, e.g. i1 on x86-64. It is also not
2057 /// necessary for non-C calling conventions. The frontend should handle this
2058 /// and include all of the necessary information.
2059 virtual MVT getTypeForExtArgOrReturn(MVT VT,
2060 ISD::NodeType /*ExtendKind*/) const {
2061 MVT MinVT = getRegisterType(MVT::i32);
2062 return VT.bitsLT(MinVT) ? MinVT : VT;
2065 /// LowerOperationWrapper - This callback is invoked by the type legalizer
2066 /// to legalize nodes with an illegal operand type but legal result types.
2067 /// It replaces the LowerOperation callback in the type Legalizer.
2068 /// The reason we can not do away with LowerOperation entirely is that
2069 /// LegalizeDAG isn't yet ready to use this callback.
2070 /// TODO: Consider merging with ReplaceNodeResults.
2072 /// The target places new result values for the node in Results (their number
2073 /// and types must exactly match those of the original return values of
2074 /// the node), or leaves Results empty, which indicates that the node is not
2075 /// to be custom lowered after all.
2076 /// The default implementation calls LowerOperation.
2077 virtual void LowerOperationWrapper(SDNode *N,
2078 SmallVectorImpl<SDValue> &Results,
2079 SelectionDAG &DAG) const;
2081 /// LowerOperation - This callback is invoked for operations that are
2082 /// unsupported by the target, which are registered to use 'custom' lowering,
2083 /// and whose defined values are all legal.
2084 /// If the target has no operations that require custom lowering, it need not
2085 /// implement this. The default implementation of this aborts.
2086 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
2088 /// ReplaceNodeResults - This callback is invoked when a node result type is
2089 /// illegal for the target, and the operation was registered to use 'custom'
2090 /// lowering for that result type. The target places new result values for
2091 /// the node in Results (their number and types must exactly match those of
2092 /// the original return values of the node), or leaves Results empty, which
2093 /// indicates that the node is not to be custom lowered after all.
2095 /// If the target has no operations that require custom lowering, it need not
2096 /// implement this. The default implementation aborts.
2097 virtual void ReplaceNodeResults(SDNode * /*N*/,
2098 SmallVectorImpl<SDValue> &/*Results*/,
2099 SelectionDAG &/*DAG*/) const {
2100 llvm_unreachable("ReplaceNodeResults not implemented for this target!");
2103 /// getTargetNodeName() - This method returns the name of a target specific
2105 virtual const char *getTargetNodeName(unsigned Opcode) const;
2107 /// createFastISel - This method returns a target specific FastISel object,
2108 /// or null if the target does not support "fast" ISel.
2109 virtual FastISel *createFastISel(FunctionLoweringInfo &,
2110 const TargetLibraryInfo *) const {
2114 //===--------------------------------------------------------------------===//
2115 // Inline Asm Support hooks
2118 /// ExpandInlineAsm - This hook allows the target to expand an inline asm
2119 /// call to be explicit llvm code if it wants to. This is useful for
2120 /// turning simple inline asms into LLVM intrinsics, which gives the
2121 /// compiler more information about the behavior of the code.
2122 virtual bool ExpandInlineAsm(CallInst *) const {
2126 enum ConstraintType {
2127 C_Register, // Constraint represents specific register(s).
2128 C_RegisterClass, // Constraint represents any of register(s) in class.
2129 C_Memory, // Memory constraint.
2130 C_Other, // Something else.
2131 C_Unknown // Unsupported constraint.
2134 enum ConstraintWeight {
2136 CW_Invalid = -1, // No match.
2137 CW_Okay = 0, // Acceptable.
2138 CW_Good = 1, // Good weight.
2139 CW_Better = 2, // Better weight.
2140 CW_Best = 3, // Best weight.
2142 // Well-known weights.
2143 CW_SpecificReg = CW_Okay, // Specific register operands.
2144 CW_Register = CW_Good, // Register operands.
2145 CW_Memory = CW_Better, // Memory operands.
2146 CW_Constant = CW_Best, // Constant operand.
2147 CW_Default = CW_Okay // Default or don't know type.
2150 /// AsmOperandInfo - This contains information for each constraint that we are
2152 struct AsmOperandInfo : public InlineAsm::ConstraintInfo {
2153 /// ConstraintCode - This contains the actual string for the code, like "m".
2154 /// TargetLowering picks the 'best' code from ConstraintInfo::Codes that
2155 /// most closely matches the operand.
2156 std::string ConstraintCode;
2158 /// ConstraintType - Information about the constraint code, e.g. Register,
2159 /// RegisterClass, Memory, Other, Unknown.
2160 TargetLowering::ConstraintType ConstraintType;
2162 /// CallOperandval - If this is the result output operand or a
2163 /// clobber, this is null, otherwise it is the incoming operand to the
2164 /// CallInst. This gets modified as the asm is processed.
2165 Value *CallOperandVal;
2167 /// ConstraintVT - The ValueType for the operand value.
2170 /// isMatchingInputConstraint - Return true of this is an input operand that
2171 /// is a matching constraint like "4".
2172 bool isMatchingInputConstraint() const;
2174 /// getMatchedOperand - If this is an input matching constraint, this method
2175 /// returns the output operand it matches.
2176 unsigned getMatchedOperand() const;
2178 /// Copy constructor for copying from an AsmOperandInfo.
2179 AsmOperandInfo(const AsmOperandInfo &info)
2180 : InlineAsm::ConstraintInfo(info),
2181 ConstraintCode(info.ConstraintCode),
2182 ConstraintType(info.ConstraintType),
2183 CallOperandVal(info.CallOperandVal),
2184 ConstraintVT(info.ConstraintVT) {
2187 /// Copy constructor for copying from a ConstraintInfo.
2188 AsmOperandInfo(const InlineAsm::ConstraintInfo &info)
2189 : InlineAsm::ConstraintInfo(info),
2190 ConstraintType(TargetLowering::C_Unknown),
2191 CallOperandVal(0), ConstraintVT(MVT::Other) {
2195 typedef std::vector<AsmOperandInfo> AsmOperandInfoVector;
2197 /// ParseConstraints - Split up the constraint string from the inline
2198 /// assembly value into the specific constraints and their prefixes,
2199 /// and also tie in the associated operand values.
2200 /// If this returns an empty vector, and if the constraint string itself
2201 /// isn't empty, there was an error parsing.
2202 virtual AsmOperandInfoVector ParseConstraints(ImmutableCallSite CS) const;
2204 /// Examine constraint type and operand type and determine a weight value.
2205 /// The operand object must already have been set up with the operand type.
2206 virtual ConstraintWeight getMultipleConstraintMatchWeight(
2207 AsmOperandInfo &info, int maIndex) const;
2209 /// Examine constraint string and operand type and determine a weight value.
2210 /// The operand object must already have been set up with the operand type.
2211 virtual ConstraintWeight getSingleConstraintMatchWeight(
2212 AsmOperandInfo &info, const char *constraint) const;
2214 /// ComputeConstraintToUse - Determines the constraint code and constraint
2215 /// type to use for the specific AsmOperandInfo, setting
2216 /// OpInfo.ConstraintCode and OpInfo.ConstraintType. If the actual operand
2217 /// being passed in is available, it can be passed in as Op, otherwise an
2218 /// empty SDValue can be passed.
2219 virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2221 SelectionDAG *DAG = 0) const;
2223 /// getConstraintType - Given a constraint, return the type of constraint it
2224 /// is for this target.
2225 virtual ConstraintType getConstraintType(const std::string &Constraint) const;
2227 /// getRegForInlineAsmConstraint - Given a physical register constraint (e.g.
2228 /// {edx}), return the register number and the register class for the
2231 /// Given a register class constraint, like 'r', if this corresponds directly
2232 /// to an LLVM register class, return a register of 0 and the register class
2235 /// This should only be used for C_Register constraints. On error,
2236 /// this returns a register number of 0 and a null register class pointer..
2237 virtual std::pair<unsigned, const TargetRegisterClass*>
2238 getRegForInlineAsmConstraint(const std::string &Constraint,
2241 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2242 /// with another that has more specific requirements based on the type of the
2243 /// corresponding operand. This returns null if there is no replacement to
2245 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
2247 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2248 /// vector. If it is invalid, don't add anything to Ops.
2249 virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
2250 std::vector<SDValue> &Ops,
2251 SelectionDAG &DAG) const;
2253 //===--------------------------------------------------------------------===//
2254 // Div utility functions
2256 SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl,
2257 SelectionDAG &DAG) const;
2258 SDValue BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization,
2259 std::vector<SDNode*> *Created) const;
2260 SDValue BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization,
2261 std::vector<SDNode*> *Created) const;
2263 //===--------------------------------------------------------------------===//
2264 // Instruction Emitting Hooks
2267 // EmitInstrWithCustomInserter - This method should be implemented by targets
2268 // that mark instructions with the 'usesCustomInserter' flag. These
2269 // instructions are special in various ways, which require special support to
2270 // insert. The specified MachineInstr is created but not inserted into any
2271 // basic blocks, and this method is called to expand it into a sequence of
2272 // instructions, potentially also creating new basic blocks and control flow.
2273 virtual MachineBasicBlock *
2274 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const;
2276 /// AdjustInstrPostInstrSelection - This method should be implemented by
2277 /// targets that mark instructions with the 'hasPostISelHook' flag. These
2278 /// instructions must be adjusted after instruction selection by target hooks.
2279 /// e.g. To fill in optional defs for ARM 's' setting instructions.
2281 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const;
2284 /// GetReturnInfo - Given an LLVM IR type and return type attributes,
2285 /// compute the return value EVTs and flags, and optionally also
2286 /// the offsets, if the return value is being lowered to memory.
2287 void GetReturnInfo(Type* ReturnType, AttributeSet attr,
2288 SmallVectorImpl<ISD::OutputArg> &Outs,
2289 const TargetLowering &TLI);
2291 } // end llvm namespace