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.
59 CMOV, // ARM conditional move instructions.
63 RBIT, // ARM bitreverse instruction
65 FTOSI, // FP to sint within a FP register.
66 FTOUI, // FP to uint within a FP register.
67 SITOF, // sint to FP within a FP register.
68 UITOF, // uint to FP within a FP register.
70 SRL_FLAG, // V,Flag = srl_flag X -> srl X, 1 + save carry out.
71 SRA_FLAG, // V,Flag = sra_flag X -> sra X, 1 + save carry out.
72 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag.
74 VMOVRRD, // double to two gprs.
75 VMOVDRR, // Two gprs to double.
77 EH_SJLJ_SETJMP, // SjLj exception handling setjmp.
78 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp.
79 EH_SJLJ_DISPATCHSETUP, // SjLj exception handling dispatch setup.
81 TC_RETURN, // Tail call return pseudo.
85 DYN_ALLOC, // Dynamic allocation on the stack.
87 MEMBARRIER, // Memory barrier (DMB)
88 MEMBARRIER_MCR, // Memory barrier (MCR)
92 VCEQ, // Vector compare equal.
93 VCEQZ, // Vector compare equal to zero.
94 VCGE, // Vector compare greater than or equal.
95 VCGEZ, // Vector compare greater than or equal to zero.
96 VCLEZ, // Vector compare less than or equal to zero.
97 VCGEU, // Vector compare unsigned greater than or equal.
98 VCGT, // Vector compare greater than.
99 VCGTZ, // Vector compare greater than zero.
100 VCLTZ, // Vector compare less than zero.
101 VCGTU, // Vector compare unsigned greater than.
102 VTST, // Vector test bits.
104 // Vector shift by immediate:
106 VSHRs, // ...right (signed)
107 VSHRu, // ...right (unsigned)
108 VSHLLs, // ...left long (signed)
109 VSHLLu, // ...left long (unsigned)
110 VSHLLi, // ...left long (with maximum shift count)
111 VSHRN, // ...right narrow
113 // Vector rounding shift by immediate:
114 VRSHRs, // ...right (signed)
115 VRSHRu, // ...right (unsigned)
116 VRSHRN, // ...right narrow
118 // Vector saturating shift by immediate:
119 VQSHLs, // ...left (signed)
120 VQSHLu, // ...left (unsigned)
121 VQSHLsu, // ...left (signed to unsigned)
122 VQSHRNs, // ...right narrow (signed)
123 VQSHRNu, // ...right narrow (unsigned)
124 VQSHRNsu, // ...right narrow (signed to unsigned)
126 // Vector saturating rounding shift by immediate:
127 VQRSHRNs, // ...right narrow (signed)
128 VQRSHRNu, // ...right narrow (unsigned)
129 VQRSHRNsu, // ...right narrow (signed to unsigned)
131 // Vector shift and insert:
135 // Vector get lane (VMOV scalar to ARM core register)
136 // (These are used for 8- and 16-bit element types only.)
137 VGETLANEu, // zero-extend vector extract element
138 VGETLANEs, // sign-extend vector extract element
140 // Vector move immediate and move negated immediate:
150 VREV64, // reverse elements within 64-bit doublewords
151 VREV32, // reverse elements within 32-bit words
152 VREV16, // reverse elements within 16-bit halfwords
153 VZIP, // zip (interleave)
154 VUZP, // unzip (deinterleave)
156 VTBL1, // 1-register shuffle with mask
157 VTBL2, // 2-register shuffle with mask
159 // Vector multiply long:
161 VMULLu, // ...unsigned
163 // Operands of the standard BUILD_VECTOR node are not legalized, which
164 // is fine if BUILD_VECTORs are always lowered to shuffles or other
165 // operations, but for ARM some BUILD_VECTORs are legal as-is and their
166 // operands need to be legalized. Define an ARM-specific version of
167 // BUILD_VECTOR for this purpose.
170 // Floating-point max and min:
177 // Vector OR with immediate
179 // Vector AND with NOT of immediate
182 // Vector bitwise select
185 // Vector load N-element structure to all lanes:
186 VLD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE,
190 // NEON loads with post-increment base updates:
202 // NEON stores with post-increment base updates:
213 /// Define some predicates that are used for node matching.
215 /// getVFPf32Imm / getVFPf64Imm - If the given fp immediate can be
216 /// materialized with a VMOV.f32 / VMOV.f64 (i.e. fconsts / fconstd)
217 /// instruction, returns its 8-bit integer representation. Otherwise,
219 int getVFPf32Imm(const APFloat &FPImm);
220 int getVFPf64Imm(const APFloat &FPImm);
221 bool isBitFieldInvertedMask(unsigned v);
224 //===--------------------------------------------------------------------===//
225 // ARMTargetLowering - ARM Implementation of the TargetLowering interface
227 class ARMTargetLowering : public TargetLowering {
229 explicit ARMTargetLowering(TargetMachine &TM);
231 virtual unsigned getJumpTableEncoding(void) const;
233 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
235 /// ReplaceNodeResults - Replace the results of node with an illegal result
236 /// type with new values built out of custom code.
238 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
239 SelectionDAG &DAG) const;
241 virtual const char *getTargetNodeName(unsigned Opcode) const;
243 virtual MachineBasicBlock *
244 EmitInstrWithCustomInserter(MachineInstr *MI,
245 MachineBasicBlock *MBB) const;
247 SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const;
248 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
250 bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const;
252 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
253 /// unaligned memory accesses. of the specified type.
254 /// FIXME: Add getOptimalMemOpType to implement memcpy with NEON?
255 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const;
257 /// isLegalAddressingMode - Return true if the addressing mode represented
258 /// by AM is legal for this target, for a load/store of the specified type.
259 virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty)const;
260 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
262 /// isLegalICmpImmediate - Return true if the specified immediate is legal
263 /// icmp immediate, that is the target has icmp instructions which can
264 /// compare a register against the immediate without having to materialize
265 /// the immediate into a register.
266 virtual bool isLegalICmpImmediate(int64_t Imm) const;
268 /// isLegalAddImmediate - Return true if the specified immediate is legal
269 /// add immediate, that is the target has add instructions which can
270 /// add a register and the immediate without having to materialize
271 /// the immediate into a register.
272 virtual bool isLegalAddImmediate(int64_t Imm) const;
274 /// getPreIndexedAddressParts - returns true by value, base pointer and
275 /// offset pointer and addressing mode by reference if the node's address
276 /// can be legally represented as pre-indexed load / store address.
277 virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
279 ISD::MemIndexedMode &AM,
280 SelectionDAG &DAG) const;
282 /// getPostIndexedAddressParts - returns true by value, base pointer and
283 /// offset pointer and addressing mode by reference if this node can be
284 /// combined with a load / store to form a post-indexed load / store.
285 virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
286 SDValue &Base, SDValue &Offset,
287 ISD::MemIndexedMode &AM,
288 SelectionDAG &DAG) const;
290 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
294 const SelectionDAG &DAG,
295 unsigned Depth) const;
298 virtual bool ExpandInlineAsm(CallInst *CI) const;
300 ConstraintType getConstraintType(const std::string &Constraint) const;
302 /// Examine constraint string and operand type and determine a weight value.
303 /// The operand object must already have been set up with the operand type.
304 ConstraintWeight getSingleConstraintMatchWeight(
305 AsmOperandInfo &info, const char *constraint) const;
307 std::pair<unsigned, const TargetRegisterClass*>
308 getRegForInlineAsmConstraint(const std::string &Constraint,
311 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
312 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
313 /// true it means one of the asm constraint of the inline asm instruction
314 /// being processed is 'm'.
315 virtual void LowerAsmOperandForConstraint(SDValue Op,
316 std::string &Constraint,
317 std::vector<SDValue> &Ops,
318 SelectionDAG &DAG) const;
320 const ARMSubtarget* getSubtarget() const {
324 /// getRegClassFor - Return the register class that should be used for the
325 /// specified value type.
326 virtual TargetRegisterClass *getRegClassFor(EVT VT) const;
328 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
329 /// be used for loads / stores from the global.
330 virtual unsigned getMaximalGlobalOffset() const;
332 /// createFastISel - This method returns a target specific FastISel object,
333 /// or null if the target does not support "fast" ISel.
334 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo) const;
336 Sched::Preference getSchedulingPreference(SDNode *N) const;
338 bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const;
339 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
341 /// isFPImmLegal - Returns true if the target can instruction select the
342 /// specified FP immediate natively. If false, the legalizer will
343 /// materialize the FP immediate as a load from a constant pool.
344 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
346 virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info,
348 unsigned Intrinsic) const;
350 std::pair<const TargetRegisterClass*, uint8_t>
351 findRepresentativeClass(EVT VT) const;
354 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
355 /// make the right decision when generating code for different targets.
356 const ARMSubtarget *Subtarget;
358 const TargetRegisterInfo *RegInfo;
360 const InstrItineraryData *Itins;
362 /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
364 unsigned ARMPCLabelIndex;
366 void addTypeForNEON(EVT VT, EVT PromotedLdStVT, EVT PromotedBitwiseVT);
367 void addDRTypeForNEON(EVT VT);
368 void addQRTypeForNEON(EVT VT);
370 typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
371 void PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
372 SDValue Chain, SDValue &Arg,
373 RegsToPassVector &RegsToPass,
374 CCValAssign &VA, CCValAssign &NextVA,
376 SmallVector<SDValue, 8> &MemOpChains,
377 ISD::ArgFlagsTy Flags) const;
378 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
379 SDValue &Root, SelectionDAG &DAG,
382 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return,
383 bool isVarArg) const;
384 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
385 DebugLoc dl, SelectionDAG &DAG,
386 const CCValAssign &VA,
387 ISD::ArgFlagsTy Flags) const;
388 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
389 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
390 SDValue LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG) const;
391 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
392 const ARMSubtarget *Subtarget) const;
393 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
394 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
395 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
396 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
397 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
398 SelectionDAG &DAG) const;
399 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
400 SelectionDAG &DAG) const;
401 SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
402 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
403 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
404 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
405 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
406 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
407 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
408 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
409 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
410 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
411 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
412 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
413 const ARMSubtarget *ST) const;
415 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
417 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
418 CallingConv::ID CallConv, bool isVarArg,
419 const SmallVectorImpl<ISD::InputArg> &Ins,
420 DebugLoc dl, SelectionDAG &DAG,
421 SmallVectorImpl<SDValue> &InVals) const;
424 LowerFormalArguments(SDValue Chain,
425 CallingConv::ID CallConv, bool isVarArg,
426 const SmallVectorImpl<ISD::InputArg> &Ins,
427 DebugLoc dl, SelectionDAG &DAG,
428 SmallVectorImpl<SDValue> &InVals) const;
430 void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
431 DebugLoc dl, SDValue &Chain, unsigned ArgOffset)
434 void computeRegArea(CCState &CCInfo, MachineFunction &MF,
435 unsigned &VARegSize, unsigned &VARegSaveSize) const;
438 LowerCall(SDValue Chain, SDValue Callee,
439 CallingConv::ID CallConv, bool isVarArg,
441 const SmallVectorImpl<ISD::OutputArg> &Outs,
442 const SmallVectorImpl<SDValue> &OutVals,
443 const SmallVectorImpl<ISD::InputArg> &Ins,
444 DebugLoc dl, SelectionDAG &DAG,
445 SmallVectorImpl<SDValue> &InVals) const;
447 /// HandleByVal - Target-specific cleanup for ByVal support.
448 virtual void HandleByVal(CCState *, unsigned &) const;
450 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
451 /// for tail call optimization. Targets which want to do tail call
452 /// optimization should implement this function.
453 bool IsEligibleForTailCallOptimization(SDValue Callee,
454 CallingConv::ID CalleeCC,
456 bool isCalleeStructRet,
457 bool isCallerStructRet,
458 const SmallVectorImpl<ISD::OutputArg> &Outs,
459 const SmallVectorImpl<SDValue> &OutVals,
460 const SmallVectorImpl<ISD::InputArg> &Ins,
461 SelectionDAG& DAG) const;
463 LowerReturn(SDValue Chain,
464 CallingConv::ID CallConv, bool isVarArg,
465 const SmallVectorImpl<ISD::OutputArg> &Outs,
466 const SmallVectorImpl<SDValue> &OutVals,
467 DebugLoc dl, SelectionDAG &DAG) const;
469 virtual bool isUsedByReturnOnly(SDNode *N) const;
471 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
473 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
474 SDValue &ARMcc, SelectionDAG &DAG, DebugLoc dl) const;
475 SDValue getVFPCmp(SDValue LHS, SDValue RHS,
476 SelectionDAG &DAG, DebugLoc dl) const;
477 SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const;
479 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const;
481 MachineBasicBlock *EmitAtomicCmpSwap(MachineInstr *MI,
482 MachineBasicBlock *BB,
483 unsigned Size) const;
484 MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
485 MachineBasicBlock *BB,
487 unsigned BinOpcode) const;
488 MachineBasicBlock * EmitAtomicBinaryMinMax(MachineInstr *MI,
489 MachineBasicBlock *BB,
492 ARMCC::CondCodes Cond) const;
494 bool RemapAddSubWithFlags(MachineInstr *MI, MachineBasicBlock *BB) const;
497 enum NEONModImmType {
505 FastISel *createFastISel(FunctionLoweringInfo &funcInfo);
509 #endif // ARMISELLOWERING_H