1 //===- X86InstrInfo.h - X86 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 X86 implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #ifndef X86INSTRUCTIONINFO_H
15 #define X86INSTRUCTIONINFO_H
17 #include "llvm/Target/TargetInstrInfo.h"
18 #include "X86RegisterInfo.h"
19 #include "llvm/ADT/IndexedMap.h"
20 #include "llvm/Target/TargetRegisterInfo.h"
23 class X86RegisterInfo;
24 class X86TargetMachine;
27 // X86 specific condition code. These correspond to X86_*_COND in
28 // X86InstrInfo.td. They must be kept in synch.
49 // X86 specific implict values used for subregister inserts.
50 // This can be used to model the fact that x86-64 by default
51 // inserts 32-bit values into 64-bit registers implicitly containing zeros.
57 // Turn condition code into conditional branch opcode.
58 unsigned GetCondBranchFromCond(CondCode CC);
60 /// GetOppositeBranchCondition - Return the inverse of the specified cond,
61 /// e.g. turning COND_E to COND_NE.
62 CondCode GetOppositeBranchCondition(X86::CondCode CC);
66 /// X86II - This namespace holds all of the target specific flags that
67 /// instruction info tracks.
71 //===------------------------------------------------------------------===//
72 // Instruction types. These are the standard/most common forms for X86
76 // PseudoFrm - This represents an instruction that is a pseudo instruction
77 // or one that has not been implemented yet. It is illegal to code generate
78 // it, but tolerated for intermediate implementation stages.
81 /// Raw - This form is for instructions that don't have any operands, so
82 /// they are just a fixed opcode value, like 'leave'.
85 /// AddRegFrm - This form is used for instructions like 'push r32' that have
86 /// their one register operand added to their opcode.
89 /// MRMDestReg - This form is used for instructions that use the Mod/RM byte
90 /// to specify a destination, which in this case is a register.
94 /// MRMDestMem - This form is used for instructions that use the Mod/RM byte
95 /// to specify a destination, which in this case is memory.
99 /// MRMSrcReg - This form is used for instructions that use the Mod/RM byte
100 /// to specify a source, which in this case is a register.
104 /// MRMSrcMem - This form is used for instructions that use the Mod/RM byte
105 /// to specify a source, which in this case is memory.
109 /// MRM[0-7][rm] - These forms are used to represent instructions that use
110 /// a Mod/RM byte, and use the middle field to hold extended opcode
111 /// information. In the intel manual these are represented as /0, /1, ...
114 // First, instructions that operate on a register r/m operand...
115 MRM0r = 16, MRM1r = 17, MRM2r = 18, MRM3r = 19, // Format /0 /1 /2 /3
116 MRM4r = 20, MRM5r = 21, MRM6r = 22, MRM7r = 23, // Format /4 /5 /6 /7
118 // Next, instructions that operate on a memory r/m operand...
119 MRM0m = 24, MRM1m = 25, MRM2m = 26, MRM3m = 27, // Format /0 /1 /2 /3
120 MRM4m = 28, MRM5m = 29, MRM6m = 30, MRM7m = 31, // Format /4 /5 /6 /7
122 // MRMInitReg - This form is used for instructions whose source and
123 // destinations are the same register.
128 //===------------------------------------------------------------------===//
131 // OpSize - Set if this instruction requires an operand size prefix (0x66),
132 // which most often indicates that the instruction operates on 16 bit data
133 // instead of 32 bit data.
136 // AsSize - Set if this instruction requires an operand size prefix (0x67),
137 // which most often indicates that the instruction address 16 bit address
138 // instead of 32 bit address (or 32 bit address in 64 bit mode).
141 //===------------------------------------------------------------------===//
142 // Op0Mask - There are several prefix bytes that are used to form two byte
143 // opcodes. These are currently 0x0F, 0xF3, and 0xD8-0xDF. This mask is
144 // used to obtain the setting of this field. If no bits in this field is
145 // set, there is no prefix byte for obtaining a multibyte opcode.
148 Op0Mask = 0xF << Op0Shift,
150 // TB - TwoByte - Set if this instruction has a two byte opcode, which
151 // starts with a 0x0F byte before the real opcode.
154 // REP - The 0xF3 prefix byte indicating repetition of the following
158 // D8-DF - These escape opcodes are used by the floating point unit. These
159 // values must remain sequential.
160 D8 = 3 << Op0Shift, D9 = 4 << Op0Shift,
161 DA = 5 << Op0Shift, DB = 6 << Op0Shift,
162 DC = 7 << Op0Shift, DD = 8 << Op0Shift,
163 DE = 9 << Op0Shift, DF = 10 << Op0Shift,
165 // XS, XD - These prefix codes are for single and double precision scalar
166 // floating point operations performed in the SSE registers.
167 XD = 11 << Op0Shift, XS = 12 << Op0Shift,
169 // T8, TA - Prefix after the 0x0F prefix.
170 T8 = 13 << Op0Shift, TA = 14 << Op0Shift,
172 //===------------------------------------------------------------------===//
173 // REX_W - REX prefixes are instruction prefixes used in 64-bit mode.
174 // They are used to specify GPRs and SSE registers, 64-bit operand size,
175 // etc. We only cares about REX.W and REX.R bits and only the former is
176 // statically determined.
179 REX_W = 1 << REXShift,
181 //===------------------------------------------------------------------===//
182 // This three-bit field describes the size of an immediate operand. Zero is
183 // unused so that we can tell if we forgot to set a value.
185 ImmMask = 7 << ImmShift,
186 Imm8 = 1 << ImmShift,
187 Imm16 = 2 << ImmShift,
188 Imm32 = 3 << ImmShift,
189 Imm64 = 4 << ImmShift,
191 //===------------------------------------------------------------------===//
192 // FP Instruction Classification... Zero is non-fp instruction.
194 // FPTypeMask - Mask for all of the FP types...
196 FPTypeMask = 7 << FPTypeShift,
198 // NotFP - The default, set for instructions that do not use FP registers.
199 NotFP = 0 << FPTypeShift,
201 // ZeroArgFP - 0 arg FP instruction which implicitly pushes ST(0), f.e. fld0
202 ZeroArgFP = 1 << FPTypeShift,
204 // OneArgFP - 1 arg FP instructions which implicitly read ST(0), such as fst
205 OneArgFP = 2 << FPTypeShift,
207 // OneArgFPRW - 1 arg FP instruction which implicitly read ST(0) and write a
208 // result back to ST(0). For example, fcos, fsqrt, etc.
210 OneArgFPRW = 3 << FPTypeShift,
212 // TwoArgFP - 2 arg FP instructions which implicitly read ST(0), and an
213 // explicit argument, storing the result to either ST(0) or the implicit
214 // argument. For example: fadd, fsub, fmul, etc...
215 TwoArgFP = 4 << FPTypeShift,
217 // CompareFP - 2 arg FP instructions which implicitly read ST(0) and an
218 // explicit argument, but have no destination. Example: fucom, fucomi, ...
219 CompareFP = 5 << FPTypeShift,
221 // CondMovFP - "2 operand" floating point conditional move instructions.
222 CondMovFP = 6 << FPTypeShift,
224 // SpecialFP - Special instruction forms. Dispatch by opcode explicitly.
225 SpecialFP = 7 << FPTypeShift,
229 LOCK = 1 << LOCKShift,
231 // Bits 20 -> 23 are unused
233 OpcodeMask = 0xFF << OpcodeShift
237 class X86InstrInfo : public TargetInstrInfoImpl {
238 X86TargetMachine &TM;
239 const X86RegisterInfo RI;
241 /// RegOp2MemOpTable2Addr, RegOp2MemOpTable0, RegOp2MemOpTable1,
242 /// RegOp2MemOpTable2 - Load / store folding opcode maps.
244 DenseMap<unsigned*, unsigned> RegOp2MemOpTable2Addr;
245 DenseMap<unsigned*, unsigned> RegOp2MemOpTable0;
246 DenseMap<unsigned*, unsigned> RegOp2MemOpTable1;
247 DenseMap<unsigned*, unsigned> RegOp2MemOpTable2;
249 /// MemOp2RegOpTable - Load / store unfolding opcode map.
251 DenseMap<unsigned*, std::pair<unsigned, unsigned> > MemOp2RegOpTable;
254 X86InstrInfo(X86TargetMachine &tm);
256 /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As
257 /// such, whenever a client has an instance of instruction info, it should
258 /// always be able to get register info as well (through this method).
260 virtual const TargetRegisterInfo &getRegisterInfo() const { return RI; }
262 // Return true if the instruction is a register to register move and
263 // leave the source and dest operands in the passed parameters.
265 bool isMoveInstr(const MachineInstr& MI, unsigned& sourceReg,
266 unsigned& destReg) const;
267 unsigned isLoadFromStackSlot(MachineInstr *MI, int &FrameIndex) const;
268 unsigned isStoreToStackSlot(MachineInstr *MI, int &FrameIndex) const;
269 bool isReallyTriviallyReMaterializable(MachineInstr *MI) const;
270 bool isInvariantLoad(MachineInstr *MI) const;
272 /// convertToThreeAddress - This method must be implemented by targets that
273 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
274 /// may be able to convert a two-address instruction into a true
275 /// three-address instruction on demand. This allows the X86 target (for
276 /// example) to convert ADD and SHL instructions into LEA instructions if they
277 /// would require register copies due to two-addressness.
279 /// This method returns a null pointer if the transformation cannot be
280 /// performed, otherwise it returns the new instruction.
282 virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
283 MachineBasicBlock::iterator &MBBI,
284 LiveVariables &LV) const;
286 /// commuteInstruction - We have a few instructions that must be hacked on to
289 virtual MachineInstr *commuteInstruction(MachineInstr *MI) const;
292 virtual bool isUnpredicatedTerminator(const MachineInstr* MI) const;
293 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
294 MachineBasicBlock *&FBB,
295 std::vector<MachineOperand> &Cond) const;
296 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const;
297 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
298 MachineBasicBlock *FBB,
299 const std::vector<MachineOperand> &Cond) const;
300 virtual void copyRegToReg(MachineBasicBlock &MBB,
301 MachineBasicBlock::iterator MI,
302 unsigned DestReg, unsigned SrcReg,
303 const TargetRegisterClass *DestRC,
304 const TargetRegisterClass *SrcRC) const;
305 virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
306 MachineBasicBlock::iterator MI,
307 unsigned SrcReg, bool isKill, int FrameIndex,
308 const TargetRegisterClass *RC) const;
310 virtual void storeRegToAddr(MachineFunction &MF, unsigned SrcReg, bool isKill,
311 SmallVectorImpl<MachineOperand> &Addr,
312 const TargetRegisterClass *RC,
313 SmallVectorImpl<MachineInstr*> &NewMIs) const;
315 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
316 MachineBasicBlock::iterator MI,
317 unsigned DestReg, int FrameIndex,
318 const TargetRegisterClass *RC) const;
320 virtual void loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
321 SmallVectorImpl<MachineOperand> &Addr,
322 const TargetRegisterClass *RC,
323 SmallVectorImpl<MachineInstr*> &NewMIs) const;
325 virtual bool spillCalleeSavedRegisters(MachineBasicBlock &MBB,
326 MachineBasicBlock::iterator MI,
327 const std::vector<CalleeSavedInfo> &CSI) const;
329 virtual bool restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
330 MachineBasicBlock::iterator MI,
331 const std::vector<CalleeSavedInfo> &CSI) const;
333 /// foldMemoryOperand - If this target supports it, fold a load or store of
334 /// the specified stack slot into the specified machine instruction for the
335 /// specified operand(s). If this is possible, the target should perform the
336 /// folding and return true, otherwise it should return false. If it folds
337 /// the instruction, it is likely that the MachineInstruction the iterator
338 /// references has been changed.
339 virtual MachineInstr* foldMemoryOperand(MachineFunction &MF,
341 SmallVectorImpl<unsigned> &Ops,
342 int FrameIndex) const;
344 /// foldMemoryOperand - Same as the previous version except it allows folding
345 /// of any load and store from / to any address, not just from a specific
347 virtual MachineInstr* foldMemoryOperand(MachineFunction &MF,
349 SmallVectorImpl<unsigned> &Ops,
350 MachineInstr* LoadMI) const;
352 /// canFoldMemoryOperand - Returns true if the specified load / store is
353 /// folding is possible.
354 virtual bool canFoldMemoryOperand(MachineInstr*, SmallVectorImpl<unsigned> &) const;
356 /// unfoldMemoryOperand - Separate a single instruction which folded a load or
357 /// a store or a load and a store into two or more instruction. If this is
358 /// possible, returns true as well as the new instructions by reference.
359 virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
360 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
361 SmallVectorImpl<MachineInstr*> &NewMIs) const;
363 virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
364 SmallVectorImpl<SDNode*> &NewNodes) const;
366 /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
367 /// instruction after load / store are unfolded from an instruction of the
368 /// specified opcode. It returns zero if the specified unfolding is not
370 virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
371 bool UnfoldLoad, bool UnfoldStore) const;
373 virtual bool BlockHasNoFallThrough(MachineBasicBlock &MBB) const;
374 virtual bool ReverseBranchCondition(std::vector<MachineOperand> &Cond) const;
376 const TargetRegisterClass *getPointerRegClass() const;
378 // getBaseOpcodeFor - This function returns the "base" X86 opcode for the
379 // specified machine instruction.
381 unsigned char getBaseOpcodeFor(const TargetInstrDesc *TID) const {
382 return TID->TSFlags >> X86II::OpcodeShift;
384 unsigned char getBaseOpcodeFor(unsigned Opcode) const {
385 return getBaseOpcodeFor(&get(Opcode));
389 MachineInstr* foldMemoryOperand(MachineInstr* MI,
391 SmallVector<MachineOperand,4> &MOs) const;
394 } // End llvm namespace