1 //===- X86InstrInfo.h - X86 Instruction Information ------------*- C++ -*- ===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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"
21 class X86RegisterInfo;
22 class X86TargetMachine;
25 // X86 specific condition code. These correspond to X86_*_COND in
26 // X86InstrInfo.td. They must be kept in synch.
47 // Turn condition code into conditional branch opcode.
48 unsigned GetCondBranchFromCond(CondCode CC);
50 /// GetOppositeBranchCondition - Return the inverse of the specified cond,
51 /// e.g. turning COND_E to COND_NE.
52 CondCode GetOppositeBranchCondition(X86::CondCode CC);
56 /// X86II - This namespace holds all of the target specific flags that
57 /// instruction info tracks.
61 //===------------------------------------------------------------------===//
62 // Instruction types. These are the standard/most common forms for X86
66 // PseudoFrm - This represents an instruction that is a pseudo instruction
67 // or one that has not been implemented yet. It is illegal to code generate
68 // it, but tolerated for intermediate implementation stages.
71 /// Raw - This form is for instructions that don't have any operands, so
72 /// they are just a fixed opcode value, like 'leave'.
75 /// AddRegFrm - This form is used for instructions like 'push r32' that have
76 /// their one register operand added to their opcode.
79 /// MRMDestReg - This form is used for instructions that use the Mod/RM byte
80 /// to specify a destination, which in this case is a register.
84 /// MRMDestMem - This form is used for instructions that use the Mod/RM byte
85 /// to specify a destination, which in this case is memory.
89 /// MRMSrcReg - This form is used for instructions that use the Mod/RM byte
90 /// to specify a source, which in this case is a register.
94 /// MRMSrcMem - This form is used for instructions that use the Mod/RM byte
95 /// to specify a source, which in this case is memory.
99 /// MRM[0-7][rm] - These forms are used to represent instructions that use
100 /// a Mod/RM byte, and use the middle field to hold extended opcode
101 /// information. In the intel manual these are represented as /0, /1, ...
104 // First, instructions that operate on a register r/m operand...
105 MRM0r = 16, MRM1r = 17, MRM2r = 18, MRM3r = 19, // Format /0 /1 /2 /3
106 MRM4r = 20, MRM5r = 21, MRM6r = 22, MRM7r = 23, // Format /4 /5 /6 /7
108 // Next, instructions that operate on a memory r/m operand...
109 MRM0m = 24, MRM1m = 25, MRM2m = 26, MRM3m = 27, // Format /0 /1 /2 /3
110 MRM4m = 28, MRM5m = 29, MRM6m = 30, MRM7m = 31, // Format /4 /5 /6 /7
112 // MRMInitReg - This form is used for instructions whose source and
113 // destinations are the same register.
118 //===------------------------------------------------------------------===//
121 // OpSize - Set if this instruction requires an operand size prefix (0x66),
122 // which most often indicates that the instruction operates on 16 bit data
123 // instead of 32 bit data.
126 // AsSize - Set if this instruction requires an operand size prefix (0x67),
127 // which most often indicates that the instruction address 16 bit address
128 // instead of 32 bit address (or 32 bit address in 64 bit mode).
131 //===------------------------------------------------------------------===//
132 // Op0Mask - There are several prefix bytes that are used to form two byte
133 // opcodes. These are currently 0x0F, 0xF3, and 0xD8-0xDF. This mask is
134 // used to obtain the setting of this field. If no bits in this field is
135 // set, there is no prefix byte for obtaining a multibyte opcode.
138 Op0Mask = 0xF << Op0Shift,
140 // TB - TwoByte - Set if this instruction has a two byte opcode, which
141 // starts with a 0x0F byte before the real opcode.
144 // REP - The 0xF3 prefix byte indicating repetition of the following
148 // D8-DF - These escape opcodes are used by the floating point unit. These
149 // values must remain sequential.
150 D8 = 3 << Op0Shift, D9 = 4 << Op0Shift,
151 DA = 5 << Op0Shift, DB = 6 << Op0Shift,
152 DC = 7 << Op0Shift, DD = 8 << Op0Shift,
153 DE = 9 << Op0Shift, DF = 10 << Op0Shift,
155 // XS, XD - These prefix codes are for single and double precision scalar
156 // floating point operations performed in the SSE registers.
157 XD = 11 << Op0Shift, XS = 12 << Op0Shift,
159 //===------------------------------------------------------------------===//
160 // REX_W - REX prefixes are instruction prefixes used in 64-bit mode.
161 // They are used to specify GPRs and SSE registers, 64-bit operand size,
162 // etc. We only cares about REX.W and REX.R bits and only the former is
163 // statically determined.
166 REX_W = 1 << REXShift,
168 //===------------------------------------------------------------------===//
169 // This three-bit field describes the size of an immediate operand. Zero is
170 // unused so that we can tell if we forgot to set a value.
172 ImmMask = 7 << ImmShift,
173 Imm8 = 1 << ImmShift,
174 Imm16 = 2 << ImmShift,
175 Imm32 = 3 << ImmShift,
176 Imm64 = 4 << ImmShift,
178 //===------------------------------------------------------------------===//
179 // FP Instruction Classification... Zero is non-fp instruction.
181 // FPTypeMask - Mask for all of the FP types...
183 FPTypeMask = 7 << FPTypeShift,
185 // NotFP - The default, set for instructions that do not use FP registers.
186 NotFP = 0 << FPTypeShift,
188 // ZeroArgFP - 0 arg FP instruction which implicitly pushes ST(0), f.e. fld0
189 ZeroArgFP = 1 << FPTypeShift,
191 // OneArgFP - 1 arg FP instructions which implicitly read ST(0), such as fst
192 OneArgFP = 2 << FPTypeShift,
194 // OneArgFPRW - 1 arg FP instruction which implicitly read ST(0) and write a
195 // result back to ST(0). For example, fcos, fsqrt, etc.
197 OneArgFPRW = 3 << FPTypeShift,
199 // TwoArgFP - 2 arg FP instructions which implicitly read ST(0), and an
200 // explicit argument, storing the result to either ST(0) or the implicit
201 // argument. For example: fadd, fsub, fmul, etc...
202 TwoArgFP = 4 << FPTypeShift,
204 // CompareFP - 2 arg FP instructions which implicitly read ST(0) and an
205 // explicit argument, but have no destination. Example: fucom, fucomi, ...
206 CompareFP = 5 << FPTypeShift,
208 // CondMovFP - "2 operand" floating point conditional move instructions.
209 CondMovFP = 6 << FPTypeShift,
211 // SpecialFP - Special instruction forms. Dispatch by opcode explicitly.
212 SpecialFP = 7 << FPTypeShift,
214 // Bits 19 -> 23 are unused
216 OpcodeMask = 0xFF << OpcodeShift
220 class X86InstrInfo : public TargetInstrInfo {
221 X86TargetMachine &TM;
222 const X86RegisterInfo RI;
224 X86InstrInfo(X86TargetMachine &tm);
226 /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As
227 /// such, whenever a client has an instance of instruction info, it should
228 /// always be able to get register info as well (through this method).
230 virtual const MRegisterInfo &getRegisterInfo() const { return RI; }
232 // Return true if the instruction is a register to register move and
233 // leave the source and dest operands in the passed parameters.
235 bool isMoveInstr(const MachineInstr& MI, unsigned& sourceReg,
236 unsigned& destReg) const;
237 unsigned isLoadFromStackSlot(MachineInstr *MI, int &FrameIndex) const;
238 unsigned isStoreToStackSlot(MachineInstr *MI, int &FrameIndex) const;
240 /// getDWARF_LABELOpcode - Return the opcode of the target's DWARF_LABEL
241 /// instruction if it has one. This is used by codegen passes that update
242 /// DWARF line number info as they modify the code.
243 virtual unsigned getDWARF_LABELOpcode() const;
245 /// convertToThreeAddress - This method must be implemented by targets that
246 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
247 /// may be able to convert a two-address instruction into a true
248 /// three-address instruction on demand. This allows the X86 target (for
249 /// example) to convert ADD and SHL instructions into LEA instructions if they
250 /// would require register copies due to two-addressness.
252 /// This method returns a null pointer if the transformation cannot be
253 /// performed, otherwise it returns the new instruction.
255 virtual MachineInstr *convertToThreeAddress(MachineInstr *TA) const;
257 /// commuteInstruction - We have a few instructions that must be hacked on to
260 virtual MachineInstr *commuteInstruction(MachineInstr *MI) const;
263 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
264 MachineBasicBlock *&FBB,
265 std::vector<MachineOperand> &Cond) const;
266 virtual void RemoveBranch(MachineBasicBlock &MBB) const;
267 virtual void InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
268 MachineBasicBlock *FBB,
269 const std::vector<MachineOperand> &Cond) const;
270 virtual bool BlockHasNoFallThrough(MachineBasicBlock &MBB) const;
271 virtual bool ReverseBranchCondition(std::vector<MachineOperand> &Cond) const;
273 const TargetRegisterClass *getPointerRegClass() const;
275 // getBaseOpcodeFor - This function returns the "base" X86 opcode for the
276 // specified opcode number.
278 unsigned char getBaseOpcodeFor(unsigned Opcode) const {
279 return get(Opcode).TSFlags >> X86II::OpcodeShift;
283 } // End llvm namespace