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
19 #include "ARMSubtarget.h"
20 #include "llvm/Target/TargetLowering.h"
21 #include "llvm/Target/TargetRegisterInfo.h"
22 #include "llvm/CodeGen/FastISel.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
28 class ARMConstantPoolValue;
31 // ARM Specific DAG Nodes
33 // Start the numbering where the builtin ops and target ops leave off.
34 FIRST_NUMBER = ISD::BUILTIN_OP_END,
36 Wrapper, // Wrapper - A wrapper node for TargetConstantPool,
37 // TargetExternalSymbol, and TargetGlobalAddress.
38 WrapperDYN, // WrapperDYN - A wrapper node for TargetGlobalAddress in
40 WrapperPIC, // WrapperPIC - A wrapper node for TargetGlobalAddress in
42 WrapperJT, // WrapperJT - A wrapper node for TargetJumpTable
44 // Add pseudo op to model memcpy for struct byval.
47 CALL, // Function call.
48 CALL_PRED, // Function call that's predicable.
49 CALL_NOLINK, // Function call with branch not branch-and-link.
50 tCALL, // Thumb function call.
51 BRCOND, // Conditional branch.
52 BR_JT, // Jumptable branch.
53 BR2_JT, // Jumptable branch (2 level - jumptable entry is a jump).
54 RET_FLAG, // Return with a flag operand.
56 PIC_ADD, // Add with a PC operand and a PIC label.
58 CMP, // ARM compare instructions.
59 CMN, // ARM CMN instructions.
60 CMPZ, // ARM compare that sets only Z flag.
61 CMPFP, // ARM VFP compare instruction, sets FPSCR.
62 CMPFPw0, // ARM VFP compare against zero instruction, sets FPSCR.
63 FMSTAT, // ARM fmstat instruction.
65 CMOV, // ARM conditional move instructions.
69 RBIT, // ARM bitreverse instruction
71 FTOSI, // FP to sint within a FP register.
72 FTOUI, // FP to uint within a FP register.
73 SITOF, // sint to FP within a FP register.
74 UITOF, // uint to FP within a FP register.
76 SRL_FLAG, // V,Flag = srl_flag X -> srl X, 1 + save carry out.
77 SRA_FLAG, // V,Flag = sra_flag X -> sra X, 1 + save carry out.
78 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag.
80 ADDC, // Add with carry
81 ADDE, // Add using carry
82 SUBC, // Sub with carry
83 SUBE, // Sub using carry
85 VMOVRRD, // double to two gprs.
86 VMOVDRR, // Two gprs to double.
88 EH_SJLJ_SETJMP, // SjLj exception handling setjmp.
89 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp.
91 TC_RETURN, // Tail call return pseudo.
95 DYN_ALLOC, // Dynamic allocation on the stack.
97 MEMBARRIER, // Memory barrier (DMB)
98 MEMBARRIER_MCR, // Memory barrier (MCR)
102 VCEQ, // Vector compare equal.
103 VCEQZ, // Vector compare equal to zero.
104 VCGE, // Vector compare greater than or equal.
105 VCGEZ, // Vector compare greater than or equal to zero.
106 VCLEZ, // Vector compare less than or equal to zero.
107 VCGEU, // Vector compare unsigned greater than or equal.
108 VCGT, // Vector compare greater than.
109 VCGTZ, // Vector compare greater than zero.
110 VCLTZ, // Vector compare less than zero.
111 VCGTU, // Vector compare unsigned greater than.
112 VTST, // Vector test bits.
114 // Vector shift by immediate:
116 VSHRs, // ...right (signed)
117 VSHRu, // ...right (unsigned)
118 VSHLLs, // ...left long (signed)
119 VSHLLu, // ...left long (unsigned)
120 VSHLLi, // ...left long (with maximum shift count)
121 VSHRN, // ...right narrow
123 // Vector rounding shift by immediate:
124 VRSHRs, // ...right (signed)
125 VRSHRu, // ...right (unsigned)
126 VRSHRN, // ...right narrow
128 // Vector saturating shift by immediate:
129 VQSHLs, // ...left (signed)
130 VQSHLu, // ...left (unsigned)
131 VQSHLsu, // ...left (signed to unsigned)
132 VQSHRNs, // ...right narrow (signed)
133 VQSHRNu, // ...right narrow (unsigned)
134 VQSHRNsu, // ...right narrow (signed to unsigned)
136 // Vector saturating rounding shift by immediate:
137 VQRSHRNs, // ...right narrow (signed)
138 VQRSHRNu, // ...right narrow (unsigned)
139 VQRSHRNsu, // ...right narrow (signed to unsigned)
141 // Vector shift and insert:
145 // Vector get lane (VMOV scalar to ARM core register)
146 // (These are used for 8- and 16-bit element types only.)
147 VGETLANEu, // zero-extend vector extract element
148 VGETLANEs, // sign-extend vector extract element
150 // Vector move immediate and move negated immediate:
154 // Vector move f32 immediate:
163 VREV64, // reverse elements within 64-bit doublewords
164 VREV32, // reverse elements within 32-bit words
165 VREV16, // reverse elements within 16-bit halfwords
166 VZIP, // zip (interleave)
167 VUZP, // unzip (deinterleave)
169 VTBL1, // 1-register shuffle with mask
170 VTBL2, // 2-register shuffle with mask
172 // Vector multiply long:
174 VMULLu, // ...unsigned
176 // Operands of the standard BUILD_VECTOR node are not legalized, which
177 // is fine if BUILD_VECTORs are always lowered to shuffles or other
178 // operations, but for ARM some BUILD_VECTORs are legal as-is and their
179 // operands need to be legalized. Define an ARM-specific version of
180 // BUILD_VECTOR for this purpose.
183 // Floating-point max and min:
190 // Vector OR with immediate
192 // Vector AND with NOT of immediate
195 // Vector bitwise select
198 // Vector load N-element structure to all lanes:
199 VLD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE,
203 // NEON loads with post-increment base updates:
215 // NEON stores with post-increment base updates:
224 // 64-bit atomic ops (value split into two registers)
236 /// Define some predicates that are used for node matching.
238 bool isBitFieldInvertedMask(unsigned v);
241 //===--------------------------------------------------------------------===//
242 // ARMTargetLowering - ARM Implementation of the TargetLowering interface
244 class ARMTargetLowering : public TargetLowering {
246 explicit ARMTargetLowering(TargetMachine &TM);
248 virtual unsigned getJumpTableEncoding(void) const;
250 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
252 /// ReplaceNodeResults - Replace the results of node with an illegal result
253 /// type with new values built out of custom code.
255 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
256 SelectionDAG &DAG) const;
258 virtual const char *getTargetNodeName(unsigned Opcode) const;
260 virtual bool isSelectSupported(SelectSupportKind Kind) const {
261 // ARM does not support scalar condition selects on vectors.
262 return (Kind != ScalarCondVectorVal);
265 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
266 virtual EVT getSetCCResultType(EVT VT) const;
268 virtual MachineBasicBlock *
269 EmitInstrWithCustomInserter(MachineInstr *MI,
270 MachineBasicBlock *MBB) const;
273 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const;
275 SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const;
276 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
278 bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const;
280 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
281 /// unaligned memory accesses. of the specified type.
282 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const;
284 virtual EVT getOptimalMemOpType(uint64_t Size,
285 unsigned DstAlign, unsigned SrcAlign,
288 MachineFunction &MF) const;
290 /// isLegalAddressingMode - Return true if the addressing mode represented
291 /// by AM is legal for this target, for a load/store of the specified type.
292 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
293 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
295 /// isLegalICmpImmediate - Return true if the specified immediate is legal
296 /// icmp immediate, that is the target has icmp instructions which can
297 /// compare a register against the immediate without having to materialize
298 /// the immediate into a register.
299 virtual bool isLegalICmpImmediate(int64_t Imm) const;
301 /// isLegalAddImmediate - Return true if the specified immediate is legal
302 /// add immediate, that is the target has add instructions which can
303 /// add a register and the immediate without having to materialize
304 /// the immediate into a register.
305 virtual bool isLegalAddImmediate(int64_t Imm) const;
307 /// getPreIndexedAddressParts - returns true by value, base pointer and
308 /// offset pointer and addressing mode by reference if the node's address
309 /// can be legally represented as pre-indexed load / store address.
310 virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
312 ISD::MemIndexedMode &AM,
313 SelectionDAG &DAG) const;
315 /// getPostIndexedAddressParts - returns true by value, base pointer and
316 /// offset pointer and addressing mode by reference if this node can be
317 /// combined with a load / store to form a post-indexed load / store.
318 virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
319 SDValue &Base, SDValue &Offset,
320 ISD::MemIndexedMode &AM,
321 SelectionDAG &DAG) const;
323 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
326 const SelectionDAG &DAG,
327 unsigned Depth) const;
330 virtual bool ExpandInlineAsm(CallInst *CI) const;
332 ConstraintType getConstraintType(const std::string &Constraint) const;
334 /// Examine constraint string and operand type and determine a weight value.
335 /// The operand object must already have been set up with the operand type.
336 ConstraintWeight getSingleConstraintMatchWeight(
337 AsmOperandInfo &info, const char *constraint) const;
339 std::pair<unsigned, const TargetRegisterClass*>
340 getRegForInlineAsmConstraint(const std::string &Constraint,
343 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
344 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
345 /// true it means one of the asm constraint of the inline asm instruction
346 /// being processed is 'm'.
347 virtual void LowerAsmOperandForConstraint(SDValue Op,
348 std::string &Constraint,
349 std::vector<SDValue> &Ops,
350 SelectionDAG &DAG) const;
352 const ARMSubtarget* getSubtarget() const {
356 /// getRegClassFor - Return the register class that should be used for the
357 /// specified value type.
358 virtual const TargetRegisterClass *getRegClassFor(EVT VT) const;
360 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
361 /// be used for loads / stores from the global.
362 virtual unsigned getMaximalGlobalOffset() const;
364 /// createFastISel - This method returns a target specific FastISel object,
365 /// or null if the target does not support "fast" ISel.
366 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
367 const TargetLibraryInfo *libInfo) const;
369 Sched::Preference getSchedulingPreference(SDNode *N) const;
371 bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const;
372 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
374 /// isFPImmLegal - Returns true if the target can instruction select the
375 /// specified FP immediate natively. If false, the legalizer will
376 /// materialize the FP immediate as a load from a constant pool.
377 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
379 virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info,
381 unsigned Intrinsic) const;
383 std::pair<const TargetRegisterClass*, uint8_t>
384 findRepresentativeClass(EVT VT) const;
387 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
388 /// make the right decision when generating code for different targets.
389 const ARMSubtarget *Subtarget;
391 const TargetRegisterInfo *RegInfo;
393 const InstrItineraryData *Itins;
395 /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
397 unsigned ARMPCLabelIndex;
399 void addTypeForNEON(MVT VT, MVT PromotedLdStVT, MVT PromotedBitwiseVT);
400 void addDRTypeForNEON(MVT VT);
401 void addQRTypeForNEON(MVT VT);
403 typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
404 void PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
405 SDValue Chain, SDValue &Arg,
406 RegsToPassVector &RegsToPass,
407 CCValAssign &VA, CCValAssign &NextVA,
409 SmallVector<SDValue, 8> &MemOpChains,
410 ISD::ArgFlagsTy Flags) const;
411 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
412 SDValue &Root, SelectionDAG &DAG,
415 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return,
416 bool isVarArg) const;
417 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
418 DebugLoc dl, SelectionDAG &DAG,
419 const CCValAssign &VA,
420 ISD::ArgFlagsTy Flags) const;
421 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
422 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
423 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
424 const ARMSubtarget *Subtarget) const;
425 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
426 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
427 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
428 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
429 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
430 SelectionDAG &DAG) const;
431 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
433 TLSModel::Model model) const;
434 SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
435 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
436 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
437 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
438 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
439 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
440 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
441 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
442 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
443 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
444 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
445 SDValue LowerConstantFP(SDValue Op, SelectionDAG &DAG,
446 const ARMSubtarget *ST) const;
447 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
448 const ARMSubtarget *ST) const;
450 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
452 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
453 CallingConv::ID CallConv, bool isVarArg,
454 const SmallVectorImpl<ISD::InputArg> &Ins,
455 DebugLoc dl, SelectionDAG &DAG,
456 SmallVectorImpl<SDValue> &InVals) const;
459 LowerFormalArguments(SDValue Chain,
460 CallingConv::ID CallConv, bool isVarArg,
461 const SmallVectorImpl<ISD::InputArg> &Ins,
462 DebugLoc dl, SelectionDAG &DAG,
463 SmallVectorImpl<SDValue> &InVals) const;
465 void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
466 DebugLoc dl, SDValue &Chain, unsigned ArgOffset)
469 void computeRegArea(CCState &CCInfo, MachineFunction &MF,
470 unsigned &VARegSize, unsigned &VARegSaveSize) const;
473 LowerCall(TargetLowering::CallLoweringInfo &CLI,
474 SmallVectorImpl<SDValue> &InVals) const;
476 /// HandleByVal - Target-specific cleanup for ByVal support.
477 virtual void HandleByVal(CCState *, unsigned &) const;
479 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
480 /// for tail call optimization. Targets which want to do tail call
481 /// optimization should implement this function.
482 bool IsEligibleForTailCallOptimization(SDValue Callee,
483 CallingConv::ID CalleeCC,
485 bool isCalleeStructRet,
486 bool isCallerStructRet,
487 const SmallVectorImpl<ISD::OutputArg> &Outs,
488 const SmallVectorImpl<SDValue> &OutVals,
489 const SmallVectorImpl<ISD::InputArg> &Ins,
490 SelectionDAG& DAG) const;
492 LowerReturn(SDValue Chain,
493 CallingConv::ID CallConv, bool isVarArg,
494 const SmallVectorImpl<ISD::OutputArg> &Outs,
495 const SmallVectorImpl<SDValue> &OutVals,
496 DebugLoc dl, SelectionDAG &DAG) const;
498 virtual bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const;
500 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
502 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
503 SDValue &ARMcc, SelectionDAG &DAG, DebugLoc dl) const;
504 SDValue getVFPCmp(SDValue LHS, SDValue RHS,
505 SelectionDAG &DAG, DebugLoc dl) const;
506 SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const;
508 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const;
510 MachineBasicBlock *EmitAtomicCmpSwap(MachineInstr *MI,
511 MachineBasicBlock *BB,
512 unsigned Size) const;
513 MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
514 MachineBasicBlock *BB,
516 unsigned BinOpcode) const;
517 MachineBasicBlock *EmitAtomicBinary64(MachineInstr *MI,
518 MachineBasicBlock *BB,
521 bool NeedsCarry = false,
522 bool IsCmpxchg = false) const;
523 MachineBasicBlock * EmitAtomicBinaryMinMax(MachineInstr *MI,
524 MachineBasicBlock *BB,
527 ARMCC::CondCodes Cond) const;
529 void SetupEntryBlockForSjLj(MachineInstr *MI,
530 MachineBasicBlock *MBB,
531 MachineBasicBlock *DispatchBB, int FI) const;
533 MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr *MI,
534 MachineBasicBlock *MBB) const;
536 bool RemapAddSubWithFlags(MachineInstr *MI, MachineBasicBlock *BB) const;
538 MachineBasicBlock *EmitStructByval(MachineInstr *MI,
539 MachineBasicBlock *MBB) const;
542 enum NEONModImmType {
550 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
551 const TargetLibraryInfo *libInfo);
555 #endif // ARMISELLOWERING_H