1 //===- AArch64InstrInfo.h - AArch64 Instruction Information -----*- 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 contains the AArch64 implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #ifndef LLVM_LIB_TARGET_AARCH64_AARCH64INSTRINFO_H
15 #define LLVM_LIB_TARGET_AARCH64_AARCH64INSTRINFO_H
18 #include "AArch64RegisterInfo.h"
19 #include "llvm/CodeGen/MachineCombinerPattern.h"
20 #include "llvm/Target/TargetInstrInfo.h"
22 #define GET_INSTRINFO_HEADER
23 #include "AArch64GenInstrInfo.inc"
27 class AArch64Subtarget;
28 class AArch64TargetMachine;
30 class AArch64InstrInfo final : public AArch64GenInstrInfo {
31 const AArch64RegisterInfo RI;
32 const AArch64Subtarget &Subtarget;
35 explicit AArch64InstrInfo(const AArch64Subtarget &STI);
37 /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As
38 /// such, whenever a client has an instance of instruction info, it should
39 /// always be able to get register info as well (through this method).
40 const AArch64RegisterInfo &getRegisterInfo() const { return RI; }
42 unsigned getInstSizeInBytes(const MachineInstr &MI) const override;
44 bool isAsCheapAsAMove(const MachineInstr &MI) const override;
46 bool isCoalescableExtInstr(const MachineInstr &MI, unsigned &SrcReg,
47 unsigned &DstReg, unsigned &SubIdx) const override;
50 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb,
51 AliasAnalysis *AA = nullptr) const override;
53 unsigned isLoadFromStackSlot(const MachineInstr &MI,
54 int &FrameIndex) const override;
55 unsigned isStoreToStackSlot(const MachineInstr &MI,
56 int &FrameIndex) const override;
58 /// Returns true if there is a shiftable register and that the shift value
60 bool hasShiftedReg(const MachineInstr &MI) const;
62 /// Returns true if there is an extendable register and that the extending
63 /// value is non-zero.
64 bool hasExtendedReg(const MachineInstr &MI) const;
66 /// \brief Does this instruction set its full destination register to zero?
67 bool isGPRZero(const MachineInstr &MI) const;
69 /// \brief Does this instruction rename a GPR without modifying bits?
70 bool isGPRCopy(const MachineInstr &MI) const;
72 /// \brief Does this instruction rename an FPR without modifying bits?
73 bool isFPRCopy(const MachineInstr &MI) const;
75 /// Return true if this is load/store scales or extends its register offset.
76 /// This refers to scaling a dynamic index as opposed to scaled immediates.
77 /// MI should be a memory op that allows scaled addressing.
78 bool isScaledAddr(const MachineInstr &MI) const;
80 /// Return true if pairing the given load or store is hinted to be
82 bool isLdStPairSuppressed(const MachineInstr &MI) const;
84 /// Return true if this is an unscaled load/store.
85 bool isUnscaledLdSt(unsigned Opc) const;
87 /// Return true if this is an unscaled load/store.
88 bool isUnscaledLdSt(MachineInstr &MI) const;
90 static bool isPairableLdStInst(const MachineInstr &MI) {
91 switch (MI.getOpcode()) {
94 // Scaled instructions.
100 case AArch64::LDRSui:
101 case AArch64::LDRDui:
102 case AArch64::LDRQui:
103 case AArch64::LDRXui:
104 case AArch64::LDRWui:
105 case AArch64::LDRSWui:
106 // Unscaled instructions.
107 case AArch64::STURSi:
108 case AArch64::STURDi:
109 case AArch64::STURQi:
110 case AArch64::STURWi:
111 case AArch64::STURXi:
112 case AArch64::LDURSi:
113 case AArch64::LDURDi:
114 case AArch64::LDURQi:
115 case AArch64::LDURWi:
116 case AArch64::LDURXi:
117 case AArch64::LDURSWi:
122 /// \brief Return the opcode that set flags when possible. The caller is
123 /// responsible for ensuring the opc has a flag setting equivalent.
124 static unsigned convertToFlagSettingOpc(unsigned Opc, bool &Is64Bit) {
127 llvm_unreachable("Opcode has no flag setting equivalent!");
129 case AArch64::ADDWri: Is64Bit = false; return AArch64::ADDSWri;
130 case AArch64::ADDWrr: Is64Bit = false; return AArch64::ADDSWrr;
131 case AArch64::ADDWrs: Is64Bit = false; return AArch64::ADDSWrs;
132 case AArch64::ADDWrx: Is64Bit = false; return AArch64::ADDSWrx;
133 case AArch64::ANDWri: Is64Bit = false; return AArch64::ANDSWri;
134 case AArch64::ANDWrr: Is64Bit = false; return AArch64::ANDSWrr;
135 case AArch64::ANDWrs: Is64Bit = false; return AArch64::ANDSWrs;
136 case AArch64::BICWrr: Is64Bit = false; return AArch64::BICSWrr;
137 case AArch64::BICWrs: Is64Bit = false; return AArch64::BICSWrs;
138 case AArch64::SUBWri: Is64Bit = false; return AArch64::SUBSWri;
139 case AArch64::SUBWrr: Is64Bit = false; return AArch64::SUBSWrr;
140 case AArch64::SUBWrs: Is64Bit = false; return AArch64::SUBSWrs;
141 case AArch64::SUBWrx: Is64Bit = false; return AArch64::SUBSWrx;
143 case AArch64::ADDXri: Is64Bit = true; return AArch64::ADDSXri;
144 case AArch64::ADDXrr: Is64Bit = true; return AArch64::ADDSXrr;
145 case AArch64::ADDXrs: Is64Bit = true; return AArch64::ADDSXrs;
146 case AArch64::ADDXrx: Is64Bit = true; return AArch64::ADDSXrx;
147 case AArch64::ANDXri: Is64Bit = true; return AArch64::ANDSXri;
148 case AArch64::ANDXrr: Is64Bit = true; return AArch64::ANDSXrr;
149 case AArch64::ANDXrs: Is64Bit = true; return AArch64::ANDSXrs;
150 case AArch64::BICXrr: Is64Bit = true; return AArch64::BICSXrr;
151 case AArch64::BICXrs: Is64Bit = true; return AArch64::BICSXrs;
152 case AArch64::SUBXri: Is64Bit = true; return AArch64::SUBSXri;
153 case AArch64::SUBXrr: Is64Bit = true; return AArch64::SUBSXrr;
154 case AArch64::SUBXrs: Is64Bit = true; return AArch64::SUBSXrs;
155 case AArch64::SUBXrx: Is64Bit = true; return AArch64::SUBSXrx;
160 /// Return true if this is a load/store that can be potentially paired/merged.
161 bool isCandidateToMergeOrPair(MachineInstr &MI) const;
163 /// Hint that pairing the given load or store is unprofitable.
164 void suppressLdStPair(MachineInstr &MI) const;
166 bool getMemOpBaseRegImmOfs(MachineInstr &LdSt, unsigned &BaseReg,
168 const TargetRegisterInfo *TRI) const override;
170 bool getMemOpBaseRegImmOfsWidth(MachineInstr &LdSt, unsigned &BaseReg,
171 int64_t &Offset, unsigned &Width,
172 const TargetRegisterInfo *TRI) const;
174 /// Return the immediate offset of the base register in a load/store \p LdSt.
175 MachineOperand &getMemOpBaseRegImmOfsOffsetOperand(MachineInstr &LdSt) const;
177 /// \brief Returns true if opcode \p Opc is a memory operation. If it is, set
178 /// \p Scale, \p Width, \p MinOffset, and \p MaxOffset accordingly.
180 /// For unscaled instructions, \p Scale is set to 1.
181 bool getMemOpInfo(unsigned Opcode, unsigned &Scale, unsigned &Width,
182 int64_t &MinOffset, int64_t &MaxOffset) const;
184 bool shouldClusterMemOps(MachineInstr &FirstLdSt, MachineInstr &SecondLdSt,
185 unsigned NumLoads) const override;
187 MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF, int FrameIx,
188 uint64_t Offset, const MDNode *Var,
190 const DebugLoc &DL) const;
191 void copyPhysRegTuple(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
192 const DebugLoc &DL, unsigned DestReg, unsigned SrcReg,
193 bool KillSrc, unsigned Opcode,
194 llvm::ArrayRef<unsigned> Indices) const;
195 void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
196 const DebugLoc &DL, unsigned DestReg, unsigned SrcReg,
197 bool KillSrc) const override;
199 void storeRegToStackSlot(MachineBasicBlock &MBB,
200 MachineBasicBlock::iterator MBBI, unsigned SrcReg,
201 bool isKill, int FrameIndex,
202 const TargetRegisterClass *RC,
203 const TargetRegisterInfo *TRI) const override;
205 void loadRegFromStackSlot(MachineBasicBlock &MBB,
206 MachineBasicBlock::iterator MBBI, unsigned DestReg,
207 int FrameIndex, const TargetRegisterClass *RC,
208 const TargetRegisterInfo *TRI) const override;
210 // This tells target independent code that it is okay to pass instructions
211 // with subreg operands to foldMemoryOperandImpl.
212 bool isSubregFoldable() const override { return true; }
214 using TargetInstrInfo::foldMemoryOperandImpl;
216 foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
217 ArrayRef<unsigned> Ops,
218 MachineBasicBlock::iterator InsertPt, int FrameIndex,
219 LiveIntervals *LIS = nullptr) const override;
221 /// \returns true if a branch from an instruction with opcode \p BranchOpc
222 /// bytes is capable of jumping to a position \p BrOffset bytes away.
223 bool isBranchOffsetInRange(unsigned BranchOpc,
224 int64_t BrOffset) const override;
226 MachineBasicBlock *getBranchDestBlock(const MachineInstr &MI) const override;
228 bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
229 MachineBasicBlock *&FBB,
230 SmallVectorImpl<MachineOperand> &Cond,
231 bool AllowModify = false) const override;
232 unsigned removeBranch(MachineBasicBlock &MBB,
233 int *BytesRemoved = nullptr) const override;
234 unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
235 MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond,
237 int *BytesAdded = nullptr) const override;
239 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const override;
240 bool canInsertSelect(const MachineBasicBlock &, ArrayRef<MachineOperand> Cond,
241 unsigned, unsigned, int &, int &, int &) const override;
242 void insertSelect(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
243 const DebugLoc &DL, unsigned DstReg,
244 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
245 unsigned FalseReg) const override;
246 void getNoop(MCInst &NopInst) const override;
248 /// analyzeCompare - For a comparison instruction, return the source registers
249 /// in SrcReg and SrcReg2, and the value it compares against in CmpValue.
250 /// Return true if the comparison instruction can be analyzed.
251 bool analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
252 unsigned &SrcReg2, int &CmpMask,
253 int &CmpValue) const override;
254 /// optimizeCompareInstr - Convert the instruction supplying the argument to
255 /// the comparison into one that sets the zero bit in the flags register.
256 bool optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
257 unsigned SrcReg2, int CmpMask, int CmpValue,
258 const MachineRegisterInfo *MRI) const override;
259 bool optimizeCondBranch(MachineInstr &MI) const override;
261 /// Return true when a code sequence can improve throughput. It
262 /// should be called only for instructions in loops.
263 /// \param Pattern - combiner pattern
264 bool isThroughputPattern(MachineCombinerPattern Pattern) const override;
265 /// Return true when there is potentially a faster code sequence
266 /// for an instruction chain ending in <Root>. All potential patterns are
267 /// listed in the <Patterns> array.
268 bool getMachineCombinerPatterns(MachineInstr &Root,
269 SmallVectorImpl<MachineCombinerPattern> &Patterns)
271 /// Return true when Inst is associative and commutative so that it can be
273 bool isAssociativeAndCommutative(const MachineInstr &Inst) const override;
274 /// When getMachineCombinerPatterns() finds patterns, this function generates
275 /// the instructions that could replace the original code sequence
276 void genAlternativeCodeSequence(
277 MachineInstr &Root, MachineCombinerPattern Pattern,
278 SmallVectorImpl<MachineInstr *> &InsInstrs,
279 SmallVectorImpl<MachineInstr *> &DelInstrs,
280 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const override;
281 /// AArch64 supports MachineCombiner.
282 bool useMachineCombiner() const override;
284 bool expandPostRAPseudo(MachineInstr &MI) const override;
286 std::pair<unsigned, unsigned>
287 decomposeMachineOperandsTargetFlags(unsigned TF) const override;
288 ArrayRef<std::pair<unsigned, const char *>>
289 getSerializableDirectMachineOperandTargetFlags() const override;
290 ArrayRef<std::pair<unsigned, const char *>>
291 getSerializableBitmaskMachineOperandTargetFlags() const override;
293 bool isFunctionSafeToOutlineFrom(MachineFunction &MF) const override;
294 unsigned getOutliningBenefit(size_t SequenceSize, size_t Occurrences,
295 bool CanBeTailCall) const override;
296 AArch64GenInstrInfo::MachineOutlinerInstrType
297 getOutliningType(MachineInstr &MI) const override;
298 void insertOutlinerEpilogue(MachineBasicBlock &MBB,
300 bool IsTailCall) const override;
301 void insertOutlinerPrologue(MachineBasicBlock &MBB,
303 bool isTailCall) const override;
304 MachineBasicBlock::iterator
305 insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
306 MachineBasicBlock::iterator &It,
308 bool IsTailCall) const override;
309 /// Returns true if the instruction has a shift by immediate that can be
310 /// executed in one cycle less.
311 bool isFalkorShiftExtFast(const MachineInstr &MI) const;
314 /// \brief Sets the offsets on outlined instructions in \p MBB which use SP
315 /// so that they will be valid post-outlining.
317 /// \param MBB A \p MachineBasicBlock in an outlined function.
318 void fixupPostOutline(MachineBasicBlock &MBB) const;
320 void instantiateCondBranch(MachineBasicBlock &MBB, const DebugLoc &DL,
321 MachineBasicBlock *TBB,
322 ArrayRef<MachineOperand> Cond) const;
323 bool substituteCmpToZero(MachineInstr &CmpInstr, unsigned SrcReg,
324 const MachineRegisterInfo *MRI) const;
327 /// emitFrameOffset - Emit instructions as needed to set DestReg to SrcReg
328 /// plus Offset. This is intended to be used from within the prolog/epilog
329 /// insertion (PEI) pass, where a virtual scratch register may be allocated
330 /// if necessary, to be replaced by the scavenger at the end of PEI.
331 void emitFrameOffset(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
332 const DebugLoc &DL, unsigned DestReg, unsigned SrcReg,
333 int Offset, const TargetInstrInfo *TII,
334 MachineInstr::MIFlag = MachineInstr::NoFlags,
335 bool SetNZCV = false);
337 /// rewriteAArch64FrameIndex - Rewrite MI to access 'Offset' bytes from the
338 /// FP. Return false if the offset could not be handled directly in MI, and
339 /// return the left-over portion by reference.
340 bool rewriteAArch64FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
341 unsigned FrameReg, int &Offset,
342 const AArch64InstrInfo *TII);
344 /// \brief Use to report the frame offset status in isAArch64FrameOffsetLegal.
345 enum AArch64FrameOffsetStatus {
346 AArch64FrameOffsetCannotUpdate = 0x0, ///< Offset cannot apply.
347 AArch64FrameOffsetIsLegal = 0x1, ///< Offset is legal.
348 AArch64FrameOffsetCanUpdate = 0x2 ///< Offset can apply, at least partly.
351 /// \brief Check if the @p Offset is a valid frame offset for @p MI.
352 /// The returned value reports the validity of the frame offset for @p MI.
353 /// It uses the values defined by AArch64FrameOffsetStatus for that.
354 /// If result == AArch64FrameOffsetCannotUpdate, @p MI cannot be updated to
356 /// If result & AArch64FrameOffsetIsLegal, @p Offset can completely be
357 /// rewriten in @p MI.
358 /// If result & AArch64FrameOffsetCanUpdate, @p Offset contains the
359 /// amount that is off the limit of the legal offset.
360 /// If set, @p OutUseUnscaledOp will contain the whether @p MI should be
361 /// turned into an unscaled operator, which opcode is in @p OutUnscaledOp.
362 /// If set, @p EmittableOffset contains the amount that can be set in @p MI
363 /// (possibly with @p OutUnscaledOp if OutUseUnscaledOp is true) and that
364 /// is a legal offset.
365 int isAArch64FrameOffsetLegal(const MachineInstr &MI, int &Offset,
366 bool *OutUseUnscaledOp = nullptr,
367 unsigned *OutUnscaledOp = nullptr,
368 int *EmittableOffset = nullptr);
370 static inline bool isUncondBranchOpcode(int Opc) { return Opc == AArch64::B; }
372 static inline bool isCondBranchOpcode(int Opc) {
389 static inline bool isIndirectBranchOpcode(int Opc) { return Opc == AArch64::BR; }
391 } // end namespace llvm