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/CodeGen/SelectionDAG.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
25 class ARMConstantPoolValue;
28 // ARM Specific DAG Nodes
30 // Start the numbering where the builtin ops and target ops leave off.
31 FIRST_NUMBER = ISD::BUILTIN_OP_END,
33 Wrapper, // Wrapper - A wrapper node for TargetConstantPool,
34 // TargetExternalSymbol, and TargetGlobalAddress.
35 WrapperJT, // WrapperJT - A wrapper node for TargetJumpTable
37 CALL, // Function call.
38 CALL_PRED, // Function call that's predicable.
39 CALL_NOLINK, // Function call with branch not branch-and-link.
40 tCALL, // Thumb function call.
41 BRCOND, // Conditional branch.
42 BR_JT, // Jumptable branch.
43 BR2_JT, // Jumptable branch (2 level - jumptable entry is a jump).
44 RET_FLAG, // Return with a flag operand.
46 PIC_ADD, // Add with a PC operand and a PIC label.
48 CMP, // ARM compare instructions.
49 CMPZ, // ARM compare that sets only Z flag.
50 CMPFP, // ARM VFP compare instruction, sets FPSCR.
51 CMPFPw0, // ARM VFP compare against zero instruction, sets FPSCR.
52 FMSTAT, // ARM fmstat instruction.
53 CMOV, // ARM conditional move instructions.
54 CNEG, // ARM conditional negate instructions.
58 RBIT, // ARM bitreverse instruction
60 FTOSI, // FP to sint within a FP register.
61 FTOUI, // FP to uint within a FP register.
62 SITOF, // sint to FP within a FP register.
63 UITOF, // uint to FP within a FP register.
65 SRL_FLAG, // V,Flag = srl_flag X -> srl X, 1 + save carry out.
66 SRA_FLAG, // V,Flag = sra_flag X -> sra X, 1 + save carry out.
67 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag.
69 VMOVRRD, // double to two gprs.
70 VMOVDRR, // Two gprs to double.
72 EH_SJLJ_SETJMP, // SjLj exception handling setjmp.
73 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp.
75 TC_RETURN, // Tail call return pseudo.
79 DYN_ALLOC, // Dynamic allocation on the stack.
81 MEMBARRIER, // Memory barrier
82 SYNCBARRIER, // Memory sync barrier
84 VCEQ, // Vector compare equal.
85 VCGE, // Vector compare greater than or equal.
86 VCGEU, // Vector compare unsigned greater than or equal.
87 VCGT, // Vector compare greater than.
88 VCGTU, // Vector compare unsigned greater than.
89 VTST, // Vector test bits.
91 // Vector shift by immediate:
93 VSHRs, // ...right (signed)
94 VSHRu, // ...right (unsigned)
95 VSHLLs, // ...left long (signed)
96 VSHLLu, // ...left long (unsigned)
97 VSHLLi, // ...left long (with maximum shift count)
98 VSHRN, // ...right narrow
100 // Vector rounding shift by immediate:
101 VRSHRs, // ...right (signed)
102 VRSHRu, // ...right (unsigned)
103 VRSHRN, // ...right narrow
105 // Vector saturating shift by immediate:
106 VQSHLs, // ...left (signed)
107 VQSHLu, // ...left (unsigned)
108 VQSHLsu, // ...left (signed to unsigned)
109 VQSHRNs, // ...right narrow (signed)
110 VQSHRNu, // ...right narrow (unsigned)
111 VQSHRNsu, // ...right narrow (signed to unsigned)
113 // Vector saturating rounding shift by immediate:
114 VQRSHRNs, // ...right narrow (signed)
115 VQRSHRNu, // ...right narrow (unsigned)
116 VQRSHRNsu, // ...right narrow (signed to unsigned)
118 // Vector shift and insert:
122 // Vector get lane (VMOV scalar to ARM core register)
123 // (These are used for 8- and 16-bit element types only.)
124 VGETLANEu, // zero-extend vector extract element
125 VGETLANEs, // sign-extend vector extract element
127 // Vector move immediate and move negated immediate:
137 VREV64, // reverse elements within 64-bit doublewords
138 VREV32, // reverse elements within 32-bit words
139 VREV16, // reverse elements within 16-bit halfwords
140 VZIP, // zip (interleave)
141 VUZP, // unzip (deinterleave)
144 // Operands of the standard BUILD_VECTOR node are not legalized, which
145 // is fine if BUILD_VECTORs are always lowered to shuffles or other
146 // operations, but for ARM some BUILD_VECTORs are legal as-is and their
147 // operands need to be legalized. Define an ARM-specific version of
148 // BUILD_VECTOR for this purpose.
151 // Floating-point max and min:
160 /// Define some predicates that are used for node matching.
162 /// getVFPf32Imm / getVFPf64Imm - If the given fp immediate can be
163 /// materialized with a VMOV.f32 / VMOV.f64 (i.e. fconsts / fconstd)
164 /// instruction, returns its 8-bit integer representation. Otherwise,
166 int getVFPf32Imm(const APFloat &FPImm);
167 int getVFPf64Imm(const APFloat &FPImm);
168 bool isBitFieldInvertedMask(unsigned v);
171 //===--------------------------------------------------------------------===//
172 // ARMTargetLowering - ARM Implementation of the TargetLowering interface
174 class ARMTargetLowering : public TargetLowering {
176 explicit ARMTargetLowering(TargetMachine &TM);
178 virtual unsigned getJumpTableEncoding(void) const;
180 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
182 /// ReplaceNodeResults - Replace the results of node with an illegal result
183 /// type with new values built out of custom code.
185 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
186 SelectionDAG &DAG) const;
188 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
190 virtual const char *getTargetNodeName(unsigned Opcode) const;
192 virtual MachineBasicBlock *
193 EmitInstrWithCustomInserter(MachineInstr *MI,
194 MachineBasicBlock *MBB) const;
196 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
197 /// unaligned memory accesses. of the specified type.
198 /// FIXME: Add getOptimalMemOpType to implement memcpy with NEON?
199 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const;
201 /// isLegalAddressingMode - Return true if the addressing mode represented
202 /// by AM is legal for this target, for a load/store of the specified type.
203 virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty)const;
204 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
206 /// isLegalICmpImmediate - Return true if the specified immediate is legal
207 /// icmp immediate, that is the target has icmp instructions which can
208 /// compare a register against the immediate without having to materialize
209 /// the immediate into a register.
210 virtual bool isLegalICmpImmediate(int64_t Imm) const;
212 /// getPreIndexedAddressParts - returns true by value, base pointer and
213 /// offset pointer and addressing mode by reference if the node's address
214 /// can be legally represented as pre-indexed load / store address.
215 virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
217 ISD::MemIndexedMode &AM,
218 SelectionDAG &DAG) const;
220 /// getPostIndexedAddressParts - returns true by value, base pointer and
221 /// offset pointer and addressing mode by reference if this node can be
222 /// combined with a load / store to form a post-indexed load / store.
223 virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
224 SDValue &Base, SDValue &Offset,
225 ISD::MemIndexedMode &AM,
226 SelectionDAG &DAG) const;
228 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
232 const SelectionDAG &DAG,
233 unsigned Depth) const;
236 ConstraintType getConstraintType(const std::string &Constraint) const;
237 std::pair<unsigned, const TargetRegisterClass*>
238 getRegForInlineAsmConstraint(const std::string &Constraint,
240 std::vector<unsigned>
241 getRegClassForInlineAsmConstraint(const std::string &Constraint,
244 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
245 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
246 /// true it means one of the asm constraint of the inline asm instruction
247 /// being processed is 'm'.
248 virtual void LowerAsmOperandForConstraint(SDValue Op,
249 char ConstraintLetter,
250 std::vector<SDValue> &Ops,
251 SelectionDAG &DAG) const;
253 const ARMSubtarget* getSubtarget() const {
257 /// getRegClassFor - Return the register class that should be used for the
258 /// specified value type.
259 virtual TargetRegisterClass *getRegClassFor(EVT VT) const;
261 /// getFunctionAlignment - Return the Log2 alignment of this function.
262 virtual unsigned getFunctionAlignment(const Function *F) const;
264 Sched::Preference getSchedulingPreference(SDNode *N) const;
266 bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const;
267 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
269 /// isFPImmLegal - Returns true if the target can instruction select the
270 /// specified FP immediate natively. If false, the legalizer will
271 /// materialize the FP immediate as a load from a constant pool.
272 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
275 std::pair<const TargetRegisterClass*, uint8_t>
276 findRepresentativeClass(EVT VT) const;
279 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
280 /// make the right decision when generating code for different targets.
281 const ARMSubtarget *Subtarget;
283 /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
285 unsigned ARMPCLabelIndex;
287 void addTypeForNEON(EVT VT, EVT PromotedLdStVT, EVT PromotedBitwiseVT);
288 void addDRTypeForNEON(EVT VT);
289 void addQRTypeForNEON(EVT VT);
291 typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
292 void PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
293 SDValue Chain, SDValue &Arg,
294 RegsToPassVector &RegsToPass,
295 CCValAssign &VA, CCValAssign &NextVA,
297 SmallVector<SDValue, 8> &MemOpChains,
298 ISD::ArgFlagsTy Flags) const;
299 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
300 SDValue &Root, SelectionDAG &DAG,
303 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return,
304 bool isVarArg) const;
305 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
306 DebugLoc dl, SelectionDAG &DAG,
307 const CCValAssign &VA,
308 ISD::ArgFlagsTy Flags) const;
309 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
310 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
311 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
312 const ARMSubtarget *Subtarget) const;
313 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
314 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
315 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
316 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
317 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
318 SelectionDAG &DAG) const;
319 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
320 SelectionDAG &DAG) const;
321 SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
322 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
323 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
324 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
325 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
326 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
327 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
328 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
329 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
330 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
332 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
333 CallingConv::ID CallConv, bool isVarArg,
334 const SmallVectorImpl<ISD::InputArg> &Ins,
335 DebugLoc dl, SelectionDAG &DAG,
336 SmallVectorImpl<SDValue> &InVals) const;
339 LowerFormalArguments(SDValue Chain,
340 CallingConv::ID CallConv, bool isVarArg,
341 const SmallVectorImpl<ISD::InputArg> &Ins,
342 DebugLoc dl, SelectionDAG &DAG,
343 SmallVectorImpl<SDValue> &InVals) const;
346 LowerCall(SDValue Chain, SDValue Callee,
347 CallingConv::ID CallConv, bool isVarArg,
349 const SmallVectorImpl<ISD::OutputArg> &Outs,
350 const SmallVectorImpl<SDValue> &OutVals,
351 const SmallVectorImpl<ISD::InputArg> &Ins,
352 DebugLoc dl, SelectionDAG &DAG,
353 SmallVectorImpl<SDValue> &InVals) const;
355 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
356 /// for tail call optimization. Targets which want to do tail call
357 /// optimization should implement this function.
358 bool IsEligibleForTailCallOptimization(SDValue Callee,
359 CallingConv::ID CalleeCC,
361 bool isCalleeStructRet,
362 bool isCallerStructRet,
363 const SmallVectorImpl<ISD::OutputArg> &Outs,
364 const SmallVectorImpl<SDValue> &OutVals,
365 const SmallVectorImpl<ISD::InputArg> &Ins,
366 SelectionDAG& DAG) const;
368 LowerReturn(SDValue Chain,
369 CallingConv::ID CallConv, bool isVarArg,
370 const SmallVectorImpl<ISD::OutputArg> &Outs,
371 const SmallVectorImpl<SDValue> &OutVals,
372 DebugLoc dl, SelectionDAG &DAG) const;
374 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
375 SDValue &ARMcc, SelectionDAG &DAG, DebugLoc dl) const;
376 SDValue getVFPCmp(SDValue LHS, SDValue RHS,
377 SelectionDAG &DAG, DebugLoc dl) const;
379 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const;
381 MachineBasicBlock *EmitAtomicCmpSwap(MachineInstr *MI,
382 MachineBasicBlock *BB,
383 unsigned Size) const;
384 MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
385 MachineBasicBlock *BB,
387 unsigned BinOpcode) const;
392 #endif // ARMISELLOWERING_H