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 "ARMSubtarget.h"
19 #include "llvm/Target/TargetLowering.h"
20 #include "llvm/Target/TargetRegisterInfo.h"
21 #include "llvm/CodeGen/FastISel.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
27 class ARMConstantPoolValue;
30 // ARM Specific DAG Nodes
32 // Start the numbering where the builtin ops and target ops leave off.
33 FIRST_NUMBER = ISD::BUILTIN_OP_END,
35 Wrapper, // Wrapper - A wrapper node for TargetConstantPool,
36 // TargetExternalSymbol, and TargetGlobalAddress.
37 WrapperDYN, // WrapperDYN - A wrapper node for TargetGlobalAddress in
39 WrapperPIC, // WrapperPIC - A wrapper node for TargetGlobalAddress in
41 WrapperJT, // WrapperJT - A wrapper node for TargetJumpTable
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.
52 PIC_ADD, // Add with a PC operand and a PIC label.
54 CMP, // ARM compare instructions.
55 CMPZ, // ARM compare that sets only Z flag.
56 CMPFP, // ARM VFP compare instruction, sets FPSCR.
57 CMPFPw0, // ARM VFP compare against zero instruction, sets FPSCR.
58 FMSTAT, // ARM fmstat instruction.
60 CMOV, // ARM conditional move instructions.
61 CAND, // ARM conditional and instructions.
62 COR, // ARM conditional or instructions.
63 CXOR, // ARM conditional xor instructions.
67 RBIT, // ARM bitreverse instruction
69 FTOSI, // FP to sint within a FP register.
70 FTOUI, // FP to uint within a FP register.
71 SITOF, // sint to FP within a FP register.
72 UITOF, // uint to FP within a FP register.
74 SRL_FLAG, // V,Flag = srl_flag X -> srl X, 1 + save carry out.
75 SRA_FLAG, // V,Flag = sra_flag X -> sra X, 1 + save carry out.
76 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag.
78 ADDC, // Add with carry
79 ADDE, // Add using carry
80 SUBC, // Sub with carry
81 SUBE, // Sub using carry
83 VMOVRRD, // double to two gprs.
84 VMOVDRR, // Two gprs to double.
86 EH_SJLJ_SETJMP, // SjLj exception handling setjmp.
87 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp.
89 TC_RETURN, // Tail call return pseudo.
93 DYN_ALLOC, // Dynamic allocation on the stack.
95 MEMBARRIER, // Memory barrier (DMB)
96 MEMBARRIER_MCR, // Memory barrier (MCR)
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)
116 VSHLLs, // ...left long (signed)
117 VSHLLu, // ...left long (unsigned)
118 VSHLLi, // ...left long (with maximum shift count)
119 VSHRN, // ...right narrow
121 // Vector rounding shift by immediate:
122 VRSHRs, // ...right (signed)
123 VRSHRu, // ...right (unsigned)
124 VRSHRN, // ...right narrow
126 // Vector saturating shift by immediate:
127 VQSHLs, // ...left (signed)
128 VQSHLu, // ...left (unsigned)
129 VQSHLsu, // ...left (signed to unsigned)
130 VQSHRNs, // ...right narrow (signed)
131 VQSHRNu, // ...right narrow (unsigned)
132 VQSHRNsu, // ...right narrow (signed to unsigned)
134 // Vector saturating rounding shift by immediate:
135 VQRSHRNs, // ...right narrow (signed)
136 VQRSHRNu, // ...right narrow (unsigned)
137 VQRSHRNsu, // ...right narrow (signed to unsigned)
139 // Vector shift and insert:
143 // Vector get lane (VMOV scalar to ARM core register)
144 // (These are used for 8- and 16-bit element types only.)
145 VGETLANEu, // zero-extend vector extract element
146 VGETLANEs, // sign-extend vector extract element
148 // Vector move immediate and move negated immediate:
152 // Vector move f32 immediate:
161 VREV64, // reverse elements within 64-bit doublewords
162 VREV32, // reverse elements within 32-bit words
163 VREV16, // reverse elements within 16-bit halfwords
164 VZIP, // zip (interleave)
165 VUZP, // unzip (deinterleave)
167 VTBL1, // 1-register shuffle with mask
168 VTBL2, // 2-register shuffle with mask
170 // Vector multiply long:
172 VMULLu, // ...unsigned
174 // Operands of the standard BUILD_VECTOR node are not legalized, which
175 // is fine if BUILD_VECTORs are always lowered to shuffles or other
176 // operations, but for ARM some BUILD_VECTORs are legal as-is and their
177 // operands need to be legalized. Define an ARM-specific version of
178 // BUILD_VECTOR for this purpose.
181 // Floating-point max and min:
188 // Vector OR with immediate
190 // Vector AND with NOT of immediate
193 // Vector bitwise select
196 // Vector load N-element structure to all lanes:
197 VLD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE,
201 // NEON loads with post-increment base updates:
213 // NEON stores with post-increment base updates:
222 // 64-bit atomic ops (value split into two registers)
234 /// Define some predicates that are used for node matching.
236 bool isBitFieldInvertedMask(unsigned v);
239 //===--------------------------------------------------------------------===//
240 // ARMTargetLowering - ARM Implementation of the TargetLowering interface
242 class ARMTargetLowering : public TargetLowering {
244 explicit ARMTargetLowering(TargetMachine &TM);
246 virtual unsigned getJumpTableEncoding(void) const;
248 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
250 /// ReplaceNodeResults - Replace the results of node with an illegal result
251 /// type with new values built out of custom code.
253 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
254 SelectionDAG &DAG) const;
256 virtual const char *getTargetNodeName(unsigned Opcode) const;
258 /// getSetCCResultType - Return the value type to use for ISD::SETCC.
259 virtual EVT getSetCCResultType(EVT VT) const;
261 virtual MachineBasicBlock *
262 EmitInstrWithCustomInserter(MachineInstr *MI,
263 MachineBasicBlock *MBB) const;
266 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const;
268 SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const;
269 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
271 bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const;
273 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
274 /// unaligned memory accesses. of the specified type.
275 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const;
277 virtual EVT getOptimalMemOpType(uint64_t Size,
278 unsigned DstAlign, unsigned SrcAlign,
281 MachineFunction &MF) const;
283 /// isLegalAddressingMode - Return true if the addressing mode represented
284 /// by AM is legal for this target, for a load/store of the specified type.
285 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
286 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
288 /// isLegalICmpImmediate - Return true if the specified immediate is legal
289 /// icmp immediate, that is the target has icmp instructions which can
290 /// compare a register against the immediate without having to materialize
291 /// the immediate into a register.
292 virtual bool isLegalICmpImmediate(int64_t Imm) const;
294 /// isLegalAddImmediate - Return true if the specified immediate is legal
295 /// add immediate, that is the target has add instructions which can
296 /// add a register and the immediate without having to materialize
297 /// the immediate into a register.
298 virtual bool isLegalAddImmediate(int64_t Imm) const;
300 /// getPreIndexedAddressParts - returns true by value, base pointer and
301 /// offset pointer and addressing mode by reference if the node's address
302 /// can be legally represented as pre-indexed load / store address.
303 virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
305 ISD::MemIndexedMode &AM,
306 SelectionDAG &DAG) const;
308 /// getPostIndexedAddressParts - returns true by value, base pointer and
309 /// offset pointer and addressing mode by reference if this node can be
310 /// combined with a load / store to form a post-indexed load / store.
311 virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
312 SDValue &Base, SDValue &Offset,
313 ISD::MemIndexedMode &AM,
314 SelectionDAG &DAG) const;
316 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
320 const SelectionDAG &DAG,
321 unsigned Depth) const;
324 virtual bool ExpandInlineAsm(CallInst *CI) const;
326 ConstraintType getConstraintType(const std::string &Constraint) const;
328 /// Examine constraint string and operand type and determine a weight value.
329 /// The operand object must already have been set up with the operand type.
330 ConstraintWeight getSingleConstraintMatchWeight(
331 AsmOperandInfo &info, const char *constraint) const;
333 std::pair<unsigned, const TargetRegisterClass*>
334 getRegForInlineAsmConstraint(const std::string &Constraint,
337 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
338 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
339 /// true it means one of the asm constraint of the inline asm instruction
340 /// being processed is 'm'.
341 virtual void LowerAsmOperandForConstraint(SDValue Op,
342 std::string &Constraint,
343 std::vector<SDValue> &Ops,
344 SelectionDAG &DAG) const;
346 const ARMSubtarget* getSubtarget() const {
350 /// getRegClassFor - Return the register class that should be used for the
351 /// specified value type.
352 virtual const TargetRegisterClass *getRegClassFor(EVT VT) const;
354 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
355 /// be used for loads / stores from the global.
356 virtual unsigned getMaximalGlobalOffset() const;
358 /// createFastISel - This method returns a target specific FastISel object,
359 /// or null if the target does not support "fast" ISel.
360 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo) const;
362 Sched::Preference getSchedulingPreference(SDNode *N) const;
364 bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const;
365 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
367 /// isFPImmLegal - Returns true if the target can instruction select the
368 /// specified FP immediate natively. If false, the legalizer will
369 /// materialize the FP immediate as a load from a constant pool.
370 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
372 virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info,
374 unsigned Intrinsic) const;
376 std::pair<const TargetRegisterClass*, uint8_t>
377 findRepresentativeClass(EVT VT) const;
380 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
381 /// make the right decision when generating code for different targets.
382 const ARMSubtarget *Subtarget;
384 const TargetRegisterInfo *RegInfo;
386 const InstrItineraryData *Itins;
388 /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
390 unsigned ARMPCLabelIndex;
392 void addTypeForNEON(EVT VT, EVT PromotedLdStVT, EVT PromotedBitwiseVT);
393 void addDRTypeForNEON(EVT VT);
394 void addQRTypeForNEON(EVT VT);
396 typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
397 void PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
398 SDValue Chain, SDValue &Arg,
399 RegsToPassVector &RegsToPass,
400 CCValAssign &VA, CCValAssign &NextVA,
402 SmallVector<SDValue, 8> &MemOpChains,
403 ISD::ArgFlagsTy Flags) const;
404 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
405 SDValue &Root, SelectionDAG &DAG,
408 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return,
409 bool isVarArg) const;
410 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
411 DebugLoc dl, SelectionDAG &DAG,
412 const CCValAssign &VA,
413 ISD::ArgFlagsTy Flags) const;
414 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
415 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
416 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
417 const ARMSubtarget *Subtarget) const;
418 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
419 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
420 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
421 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
422 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
423 SelectionDAG &DAG) const;
424 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
425 SelectionDAG &DAG) const;
426 SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
427 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
428 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
429 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
430 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
431 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
432 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
433 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
434 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
435 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
436 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
437 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
438 const ARMSubtarget *ST) const;
440 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
442 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
443 CallingConv::ID CallConv, bool isVarArg,
444 const SmallVectorImpl<ISD::InputArg> &Ins,
445 DebugLoc dl, SelectionDAG &DAG,
446 SmallVectorImpl<SDValue> &InVals) const;
449 LowerFormalArguments(SDValue Chain,
450 CallingConv::ID CallConv, bool isVarArg,
451 const SmallVectorImpl<ISD::InputArg> &Ins,
452 DebugLoc dl, SelectionDAG &DAG,
453 SmallVectorImpl<SDValue> &InVals) const;
455 void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
456 DebugLoc dl, SDValue &Chain, unsigned ArgOffset)
459 void computeRegArea(CCState &CCInfo, MachineFunction &MF,
460 unsigned &VARegSize, unsigned &VARegSaveSize) const;
463 LowerCall(SDValue Chain, SDValue Callee,
464 CallingConv::ID CallConv, bool isVarArg,
466 const SmallVectorImpl<ISD::OutputArg> &Outs,
467 const SmallVectorImpl<SDValue> &OutVals,
468 const SmallVectorImpl<ISD::InputArg> &Ins,
469 DebugLoc dl, SelectionDAG &DAG,
470 SmallVectorImpl<SDValue> &InVals) const;
472 /// HandleByVal - Target-specific cleanup for ByVal support.
473 virtual void HandleByVal(CCState *, unsigned &) const;
475 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
476 /// for tail call optimization. Targets which want to do tail call
477 /// optimization should implement this function.
478 bool IsEligibleForTailCallOptimization(SDValue Callee,
479 CallingConv::ID CalleeCC,
481 bool isCalleeStructRet,
482 bool isCallerStructRet,
483 const SmallVectorImpl<ISD::OutputArg> &Outs,
484 const SmallVectorImpl<SDValue> &OutVals,
485 const SmallVectorImpl<ISD::InputArg> &Ins,
486 SelectionDAG& DAG) const;
488 LowerReturn(SDValue Chain,
489 CallingConv::ID CallConv, bool isVarArg,
490 const SmallVectorImpl<ISD::OutputArg> &Outs,
491 const SmallVectorImpl<SDValue> &OutVals,
492 DebugLoc dl, SelectionDAG &DAG) const;
494 virtual bool isUsedByReturnOnly(SDNode *N) const;
496 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
498 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
499 SDValue &ARMcc, SelectionDAG &DAG, DebugLoc dl) const;
500 SDValue getVFPCmp(SDValue LHS, SDValue RHS,
501 SelectionDAG &DAG, DebugLoc dl) const;
502 SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const;
504 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const;
506 MachineBasicBlock *EmitAtomicCmpSwap(MachineInstr *MI,
507 MachineBasicBlock *BB,
508 unsigned Size) const;
509 MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
510 MachineBasicBlock *BB,
512 unsigned BinOpcode) const;
513 MachineBasicBlock *EmitAtomicBinary64(MachineInstr *MI,
514 MachineBasicBlock *BB,
517 bool NeedsCarry = false,
518 bool IsCmpxchg = false) const;
519 MachineBasicBlock * EmitAtomicBinaryMinMax(MachineInstr *MI,
520 MachineBasicBlock *BB,
523 ARMCC::CondCodes Cond) const;
525 void SetupEntryBlockForSjLj(MachineInstr *MI,
526 MachineBasicBlock *MBB,
527 MachineBasicBlock *DispatchBB, int FI) const;
529 MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr *MI,
530 MachineBasicBlock *MBB) const;
532 bool RemapAddSubWithFlags(MachineInstr *MI, MachineBasicBlock *BB) const;
535 enum NEONModImmType {
543 FastISel *createFastISel(FunctionLoweringInfo &funcInfo);
547 #endif // ARMISELLOWERING_H