1 //===-- ARMISelLowering.h - ARM DAG Lowering Interface ----------*- 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 defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #ifndef ARMISELLOWERING_H
16 #define ARMISELLOWERING_H
18 #include "MCTargetDesc/ARMBaseInfo.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/Target/TargetLowering.h"
25 class ARMConstantPoolValue;
29 // ARM Specific DAG Nodes
31 // Start the numbering where the builtin ops and target ops leave off.
32 FIRST_NUMBER = ISD::BUILTIN_OP_END,
34 Wrapper, // Wrapper - A wrapper node for TargetConstantPool,
35 // TargetExternalSymbol, and TargetGlobalAddress.
36 WrapperPIC, // WrapperPIC - A wrapper node for TargetGlobalAddress in
38 WrapperJT, // WrapperJT - A wrapper node for TargetJumpTable
40 // Add pseudo op to model memcpy for struct byval.
43 CALL, // Function call.
44 CALL_PRED, // Function call that's predicable.
45 CALL_NOLINK, // Function call with branch not branch-and-link.
46 tCALL, // Thumb function call.
47 BRCOND, // Conditional branch.
48 BR_JT, // Jumptable branch.
49 BR2_JT, // Jumptable branch (2 level - jumptable entry is a jump).
50 RET_FLAG, // Return with a flag operand.
51 INTRET_FLAG, // Interrupt return with an LR-offset and a flag operand.
53 PIC_ADD, // Add with a PC operand and a PIC label.
55 CMP, // ARM compare instructions.
56 CMN, // ARM CMN instructions.
57 CMPZ, // ARM compare that sets only Z flag.
58 CMPFP, // ARM VFP compare instruction, sets FPSCR.
59 CMPFPw0, // ARM VFP compare against zero instruction, sets FPSCR.
60 FMSTAT, // ARM fmstat instruction.
62 CMOV, // ARM conditional move instructions.
66 RBIT, // ARM bitreverse instruction
68 FTOSI, // FP to sint within a FP register.
69 FTOUI, // FP to uint within a FP register.
70 SITOF, // sint to FP within a FP register.
71 UITOF, // uint to FP within a FP register.
73 SRL_FLAG, // V,Flag = srl_flag X -> srl X, 1 + save carry out.
74 SRA_FLAG, // V,Flag = sra_flag X -> sra X, 1 + save carry out.
75 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag.
77 ADDC, // Add with carry
78 ADDE, // Add using carry
79 SUBC, // Sub with carry
80 SUBE, // Sub using carry
82 VMOVRRD, // double to two gprs.
83 VMOVDRR, // Two gprs to double.
85 EH_SJLJ_SETJMP, // SjLj exception handling setjmp.
86 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp.
88 TC_RETURN, // Tail call return pseudo.
92 DYN_ALLOC, // Dynamic allocation on the stack.
94 MEMBARRIER_MCR, // Memory barrier (MCR)
98 WIN__CHKSTK, // Windows' __chkstk call to do stack probing.
100 VCEQ, // Vector compare equal.
101 VCEQZ, // Vector compare equal to zero.
102 VCGE, // Vector compare greater than or equal.
103 VCGEZ, // Vector compare greater than or equal to zero.
104 VCLEZ, // Vector compare less than or equal to zero.
105 VCGEU, // Vector compare unsigned greater than or equal.
106 VCGT, // Vector compare greater than.
107 VCGTZ, // Vector compare greater than zero.
108 VCLTZ, // Vector compare less than zero.
109 VCGTU, // Vector compare unsigned greater than.
110 VTST, // Vector test bits.
112 // Vector shift by immediate:
114 VSHRs, // ...right (signed)
115 VSHRu, // ...right (unsigned)
117 // Vector rounding shift by immediate:
118 VRSHRs, // ...right (signed)
119 VRSHRu, // ...right (unsigned)
120 VRSHRN, // ...right narrow
122 // Vector saturating shift by immediate:
123 VQSHLs, // ...left (signed)
124 VQSHLu, // ...left (unsigned)
125 VQSHLsu, // ...left (signed to unsigned)
126 VQSHRNs, // ...right narrow (signed)
127 VQSHRNu, // ...right narrow (unsigned)
128 VQSHRNsu, // ...right narrow (signed to unsigned)
130 // Vector saturating rounding shift by immediate:
131 VQRSHRNs, // ...right narrow (signed)
132 VQRSHRNu, // ...right narrow (unsigned)
133 VQRSHRNsu, // ...right narrow (signed to unsigned)
135 // Vector shift and insert:
139 // Vector get lane (VMOV scalar to ARM core register)
140 // (These are used for 8- and 16-bit element types only.)
141 VGETLANEu, // zero-extend vector extract element
142 VGETLANEs, // sign-extend vector extract element
144 // Vector move immediate and move negated immediate:
148 // Vector move f32 immediate:
157 VREV64, // reverse elements within 64-bit doublewords
158 VREV32, // reverse elements within 32-bit words
159 VREV16, // reverse elements within 16-bit halfwords
160 VZIP, // zip (interleave)
161 VUZP, // unzip (deinterleave)
163 VTBL1, // 1-register shuffle with mask
164 VTBL2, // 2-register shuffle with mask
166 // Vector multiply long:
168 VMULLu, // ...unsigned
170 UMLAL, // 64bit Unsigned Accumulate Multiply
171 SMLAL, // 64bit Signed Accumulate Multiply
173 // Operands of the standard BUILD_VECTOR node are not legalized, which
174 // is fine if BUILD_VECTORs are always lowered to shuffles or other
175 // operations, but for ARM some BUILD_VECTORs are legal as-is and their
176 // operands need to be legalized. Define an ARM-specific version of
177 // BUILD_VECTOR for this purpose.
180 // Floating-point max and min:
189 // Vector OR with immediate
191 // Vector AND with NOT of immediate
194 // Vector bitwise select
197 // Vector load N-element structure to all lanes:
198 VLD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE,
202 // NEON loads with post-increment base updates:
214 // NEON stores with post-increment base updates:
225 /// Define some predicates that are used for node matching.
227 bool isBitFieldInvertedMask(unsigned v);
230 //===--------------------------------------------------------------------===//
231 // ARMTargetLowering - ARM Implementation of the TargetLowering interface
233 class ARMTargetLowering : public TargetLowering {
235 explicit ARMTargetLowering(TargetMachine &TM);
237 unsigned getJumpTableEncoding() const override;
239 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
241 /// ReplaceNodeResults - Replace the results of node with an illegal result
242 /// type with new values built out of custom code.
244 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
245 SelectionDAG &DAG) const override;
247 const char *getTargetNodeName(unsigned Opcode) const override;
249 bool isSelectSupported(SelectSupportKind Kind) const override {
250 // ARM does not support scalar condition selects on vectors.
251 return (Kind != ScalarCondVectorVal);
254 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
255 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
258 EmitInstrWithCustomInserter(MachineInstr *MI,
259 MachineBasicBlock *MBB) const override;
261 void AdjustInstrPostInstrSelection(MachineInstr *MI,
262 SDNode *Node) const override;
264 SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const;
265 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
267 bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const override;
269 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
270 /// unaligned memory accesses of the specified type. Returns whether it
271 /// is "fast" by reference in the second argument.
272 bool allowsUnalignedMemoryAccesses(EVT VT, unsigned AddrSpace,
273 bool *Fast) const override;
275 EVT getOptimalMemOpType(uint64_t Size,
276 unsigned DstAlign, unsigned SrcAlign,
277 bool IsMemset, bool ZeroMemset,
279 MachineFunction &MF) const override;
281 using TargetLowering::isZExtFree;
282 bool isZExtFree(SDValue Val, EVT VT2) const override;
284 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override;
287 /// isLegalAddressingMode - Return true if the addressing mode represented
288 /// by AM is legal for this target, for a load/store of the specified type.
289 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
290 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
292 /// isLegalICmpImmediate - Return true if the specified immediate is legal
293 /// icmp immediate, that is the target has icmp instructions which can
294 /// compare a register against the immediate without having to materialize
295 /// the immediate into a register.
296 bool isLegalICmpImmediate(int64_t Imm) const override;
298 /// isLegalAddImmediate - Return true if the specified immediate is legal
299 /// add immediate, that is the target has add instructions which can
300 /// add a register and the immediate without having to materialize
301 /// the immediate into a register.
302 bool isLegalAddImmediate(int64_t Imm) const override;
304 /// getPreIndexedAddressParts - returns true by value, base pointer and
305 /// offset pointer and addressing mode by reference if the node's address
306 /// can be legally represented as pre-indexed load / store address.
307 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset,
308 ISD::MemIndexedMode &AM,
309 SelectionDAG &DAG) const override;
311 /// getPostIndexedAddressParts - returns true by value, base pointer and
312 /// offset pointer and addressing mode by reference if this node can be
313 /// combined with a load / store to form a post-indexed load / store.
314 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base,
315 SDValue &Offset, ISD::MemIndexedMode &AM,
316 SelectionDAG &DAG) const override;
318 void computeKnownBitsForTargetNode(const SDValue Op, APInt &KnownZero,
320 const SelectionDAG &DAG,
321 unsigned Depth) const override;
324 bool ExpandInlineAsm(CallInst *CI) const override;
327 getConstraintType(const std::string &Constraint) const override;
329 /// Examine constraint string and operand type and determine a weight value.
330 /// The operand object must already have been set up with the operand type.
331 ConstraintWeight getSingleConstraintMatchWeight(
332 AsmOperandInfo &info, const char *constraint) const override;
334 std::pair<unsigned, const TargetRegisterClass*>
335 getRegForInlineAsmConstraint(const std::string &Constraint,
336 MVT VT) const override;
338 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
339 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
340 /// true it means one of the asm constraint of the inline asm instruction
341 /// being processed is 'm'.
342 void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
343 std::vector<SDValue> &Ops,
344 SelectionDAG &DAG) const override;
346 const ARMSubtarget* getSubtarget() const {
350 /// getRegClassFor - Return the register class that should be used for the
351 /// specified value type.
352 const TargetRegisterClass *getRegClassFor(MVT VT) const override;
354 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
355 /// be used for loads / stores from the global.
356 unsigned getMaximalGlobalOffset() const override;
358 /// Returns true if a cast between SrcAS and DestAS is a noop.
359 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
360 // Addrspacecasts are always noops.
364 /// createFastISel - This method returns a target specific FastISel object,
365 /// or null if the target does not support "fast" ISel.
366 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
367 const TargetLibraryInfo *libInfo) const override;
369 Sched::Preference getSchedulingPreference(SDNode *N) const override;
372 isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const override;
373 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
375 /// isFPImmLegal - Returns true if the target can instruction select the
376 /// specified FP immediate natively. If false, the legalizer will
377 /// materialize the FP immediate as a load from a constant pool.
378 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
380 bool getTgtMemIntrinsic(IntrinsicInfo &Info,
382 unsigned Intrinsic) const override;
384 /// \brief Returns true if it is beneficial to convert a load of a constant
385 /// to just the constant itself.
386 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
387 Type *Ty) const override;
389 /// \brief Returns true if an argument of type Ty needs to be passed in a
390 /// contiguous block of registers in calling convention CallConv.
391 bool functionArgumentNeedsConsecutiveRegisters(
392 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override;
394 Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
395 AtomicOrdering Ord) const override;
396 Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val,
397 Value *Addr, AtomicOrdering Ord) const override;
399 bool shouldExpandAtomicInIR(Instruction *Inst) const override;
402 std::pair<const TargetRegisterClass*, uint8_t>
403 findRepresentativeClass(MVT VT) const override;
406 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
407 /// make the right decision when generating code for different targets.
408 const ARMSubtarget *Subtarget;
410 const TargetRegisterInfo *RegInfo;
412 const InstrItineraryData *Itins;
414 /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
416 unsigned ARMPCLabelIndex;
418 void addTypeForNEON(MVT VT, MVT PromotedLdStVT, MVT PromotedBitwiseVT);
419 void addDRTypeForNEON(MVT VT);
420 void addQRTypeForNEON(MVT VT);
421 std::pair<SDValue, SDValue> getARMXALUOOp(SDValue Op, SelectionDAG &DAG, SDValue &ARMcc) const;
423 typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
424 void PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
425 SDValue Chain, SDValue &Arg,
426 RegsToPassVector &RegsToPass,
427 CCValAssign &VA, CCValAssign &NextVA,
429 SmallVectorImpl<SDValue> &MemOpChains,
430 ISD::ArgFlagsTy Flags) const;
431 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
432 SDValue &Root, SelectionDAG &DAG,
435 CallingConv::ID getEffectiveCallingConv(CallingConv::ID CC,
436 bool isVarArg) const;
437 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return,
438 bool isVarArg) const;
439 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
440 SDLoc dl, SelectionDAG &DAG,
441 const CCValAssign &VA,
442 ISD::ArgFlagsTy Flags) const;
443 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
444 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
445 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
446 const ARMSubtarget *Subtarget) const;
447 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
448 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
449 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
450 SDValue LowerGlobalAddressWindows(SDValue Op, SelectionDAG &DAG) const;
451 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
452 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
453 SelectionDAG &DAG) const;
454 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
456 TLSModel::Model model) const;
457 SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
458 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
459 SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) const;
460 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
461 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
462 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
463 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
464 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
465 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
466 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
467 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
468 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
469 SDValue LowerConstantFP(SDValue Op, SelectionDAG &DAG,
470 const ARMSubtarget *ST) const;
471 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
472 const ARMSubtarget *ST) const;
473 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
474 SDValue LowerDivRem(SDValue Op, SelectionDAG &DAG) const;
475 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
477 unsigned getRegisterByName(const char* RegName, EVT VT) const override;
479 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
480 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
481 /// expanded to FMAs when this method returns true, otherwise fmuladd is
482 /// expanded to fmul + fadd.
484 /// ARM supports both fused and unfused multiply-add operations; we already
485 /// lower a pair of fmul and fadd to the latter so it's not clear that there
486 /// would be a gain or that the gain would be worthwhile enough to risk
487 /// correctness bugs.
488 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override { return false; }
490 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
492 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
493 CallingConv::ID CallConv, bool isVarArg,
494 const SmallVectorImpl<ISD::InputArg> &Ins,
495 SDLoc dl, SelectionDAG &DAG,
496 SmallVectorImpl<SDValue> &InVals,
497 bool isThisReturn, SDValue ThisVal) const;
500 LowerFormalArguments(SDValue Chain,
501 CallingConv::ID CallConv, bool isVarArg,
502 const SmallVectorImpl<ISD::InputArg> &Ins,
503 SDLoc dl, SelectionDAG &DAG,
504 SmallVectorImpl<SDValue> &InVals) const override;
506 int StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
507 SDLoc dl, SDValue &Chain,
508 const Value *OrigArg,
509 unsigned InRegsParamRecordIdx,
510 unsigned OffsetFromOrigArg,
514 unsigned ByValStoreOffset,
515 unsigned TotalArgRegsSaveSize) const;
517 void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
518 SDLoc dl, SDValue &Chain,
520 unsigned TotalArgRegsSaveSize,
521 bool ForceMutable = false) const;
523 void computeRegArea(CCState &CCInfo, MachineFunction &MF,
524 unsigned InRegsParamRecordIdx,
526 unsigned &ArgRegsSize,
527 unsigned &ArgRegsSaveSize) const;
530 LowerCall(TargetLowering::CallLoweringInfo &CLI,
531 SmallVectorImpl<SDValue> &InVals) const override;
533 /// HandleByVal - Target-specific cleanup for ByVal support.
534 void HandleByVal(CCState *, unsigned &, unsigned) const override;
536 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
537 /// for tail call optimization. Targets which want to do tail call
538 /// optimization should implement this function.
539 bool IsEligibleForTailCallOptimization(SDValue Callee,
540 CallingConv::ID CalleeCC,
542 bool isCalleeStructRet,
543 bool isCallerStructRet,
544 const SmallVectorImpl<ISD::OutputArg> &Outs,
545 const SmallVectorImpl<SDValue> &OutVals,
546 const SmallVectorImpl<ISD::InputArg> &Ins,
547 SelectionDAG& DAG) const;
549 bool CanLowerReturn(CallingConv::ID CallConv,
550 MachineFunction &MF, bool isVarArg,
551 const SmallVectorImpl<ISD::OutputArg> &Outs,
552 LLVMContext &Context) const override;
555 LowerReturn(SDValue Chain,
556 CallingConv::ID CallConv, bool isVarArg,
557 const SmallVectorImpl<ISD::OutputArg> &Outs,
558 const SmallVectorImpl<SDValue> &OutVals,
559 SDLoc dl, SelectionDAG &DAG) const override;
561 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
563 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
565 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
566 SDValue &ARMcc, SelectionDAG &DAG, SDLoc dl) const;
567 SDValue getVFPCmp(SDValue LHS, SDValue RHS,
568 SelectionDAG &DAG, SDLoc dl) const;
569 SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const;
571 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const;
573 void SetupEntryBlockForSjLj(MachineInstr *MI,
574 MachineBasicBlock *MBB,
575 MachineBasicBlock *DispatchBB, int FI) const;
577 MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr *MI,
578 MachineBasicBlock *MBB) const;
580 bool RemapAddSubWithFlags(MachineInstr *MI, MachineBasicBlock *BB) const;
582 MachineBasicBlock *EmitStructByval(MachineInstr *MI,
583 MachineBasicBlock *MBB) const;
585 MachineBasicBlock *EmitLowered__chkstk(MachineInstr *MI,
586 MachineBasicBlock *MBB) const;
589 enum NEONModImmType {
596 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
597 const TargetLibraryInfo *libInfo);
601 #endif // ARMISELLOWERING_H