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"
19 #include "X86RegisterInfo.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/Target/TargetRegisterInfo.h"
24 class X86RegisterInfo;
25 class X86TargetMachine;
28 // X86 specific condition code. These correspond to X86_*_COND in
29 // X86InstrInfo.td. They must be kept in synch.
48 // Artificial condition codes. These are used by AnalyzeBranch
49 // to indicate a block terminated with two conditional branches to
50 // the same location. This occurs in code using FCMP_OEQ or FCMP_UNE,
51 // which can't be represented on x86 with a single condition. These
52 // are never used in MachineInstrs.
59 // Turn condition code into conditional branch opcode.
60 unsigned GetCondBranchFromCond(CondCode CC);
62 /// GetOppositeBranchCondition - Return the inverse of the specified cond,
63 /// e.g. turning COND_E to COND_NE.
64 CondCode GetOppositeBranchCondition(X86::CondCode CC);
68 /// X86II - This namespace holds all of the target specific flags that
69 /// instruction info tracks.
73 //===------------------------------------------------------------------===//
74 // X86 Specific MachineOperand flags.
78 /// MO_GOT_ABSOLUTE_ADDRESS - On a symbol operand, this represents a
80 /// SYMBOL_LABEL + [. - PICBASELABEL]
81 MO_GOT_ABSOLUTE_ADDRESS = 1,
83 /// MO_PIC_BASE_OFFSET - On a symbol operand this indicates that the
84 /// immediate should get the value of the symbol minus the PIC base label:
85 /// SYMBOL_LABEL - PICBASELABEL
86 MO_PIC_BASE_OFFSET = 2,
88 /// MO_GOT - On a symbol operand this indicates that the immediate is the
89 /// offset to the GOT entry for the symbol name from the base of the GOT.
91 /// See the X86-64 ELF ABI supplement for more details.
95 /// MO_GOTOFF - On a symbol operand this indicates that the immediate is
96 /// the offset to the location of the symbol name from the base of the GOT.
98 /// See the X86-64 ELF ABI supplement for more details.
99 /// SYMBOL_LABEL @GOTOFF
102 /// MO_GOTPCREL - On a symbol operand this indicates that the immediate is
103 /// offset to the GOT entry for the symbol name from the current code
106 /// See the X86-64 ELF ABI supplement for more details.
107 /// SYMBOL_LABEL @GOTPCREL
110 /// MO_PLT - On a symbol operand this indicates that the immediate is
111 /// offset to the PLT entry of symbol name from the current code location.
113 /// See the X86-64 ELF ABI supplement for more details.
114 /// SYMBOL_LABEL @PLT
117 /// MO_TLSGD - On a symbol operand this indicates that the immediate is
120 /// See 'ELF Handling for Thread-Local Storage' for more details.
121 /// SYMBOL_LABEL @TLSGD
124 /// MO_GOTTPOFF - On a symbol operand this indicates that the immediate is
127 /// See 'ELF Handling for Thread-Local Storage' for more details.
128 /// SYMBOL_LABEL @GOTTPOFF
131 /// MO_INDNTPOFF - On a symbol operand this indicates that the immediate is
134 /// See 'ELF Handling for Thread-Local Storage' for more details.
135 /// SYMBOL_LABEL @INDNTPOFF
138 /// MO_TPOFF - On a symbol operand this indicates that the immediate is
141 /// See 'ELF Handling for Thread-Local Storage' for more details.
142 /// SYMBOL_LABEL @TPOFF
145 /// MO_NTPOFF - On a symbol operand this indicates that the immediate is
148 /// See 'ELF Handling for Thread-Local Storage' for more details.
149 /// SYMBOL_LABEL @NTPOFF
152 /// MO_DLLIMPORT - On a symbol operand "FOO", this indicates that the
153 /// reference is actually to the "__imp_FOO" symbol. This is used for
154 /// dllimport linkage on windows.
157 /// MO_DARWIN_STUB - On a symbol operand "FOO", this indicates that the
158 /// reference is actually to the "FOO$stub" symbol. This is used for calls
159 /// and jumps to external functions on Tiger and before.
162 /// MO_DARWIN_NONLAZY - On a symbol operand "FOO", this indicates that the
163 /// reference is actually to the "FOO$non_lazy_ptr" symbol, which is a
164 /// non-PIC-base-relative reference to a non-hidden dyld lazy pointer stub.
165 MO_DARWIN_NONLAZY = 14,
167 /// MO_DARWIN_NONLAZY_PIC_BASE - On a symbol operand "FOO", this indicates
168 /// that the reference is actually to "FOO$non_lazy_ptr - PICBASE", which is
169 /// a PIC-base-relative reference to a non-hidden dyld lazy pointer stub.
170 MO_DARWIN_NONLAZY_PIC_BASE = 15,
172 /// MO_DARWIN_HIDDEN_NONLAZY - On a symbol operand "FOO", this indicates
173 /// that the reference is actually to the "FOO$non_lazy_ptr" symbol, which
174 /// is a non-PIC-base-relative reference to a hidden dyld lazy pointer stub.
175 MO_DARWIN_HIDDEN_NONLAZY = 16,
177 /// MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE - On a symbol operand "FOO", this
178 /// indicates that the reference is actually to "FOO$non_lazy_ptr -PICBASE",
179 /// which is a PIC-base-relative reference to a hidden dyld lazy pointer
181 MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE = 17,
183 //===------------------------------------------------------------------===//
184 // Instruction encodings. These are the standard/most common forms for X86
188 // PseudoFrm - This represents an instruction that is a pseudo instruction
189 // or one that has not been implemented yet. It is illegal to code generate
190 // it, but tolerated for intermediate implementation stages.
193 /// Raw - This form is for instructions that don't have any operands, so
194 /// they are just a fixed opcode value, like 'leave'.
197 /// AddRegFrm - This form is used for instructions like 'push r32' that have
198 /// their one register operand added to their opcode.
201 /// MRMDestReg - This form is used for instructions that use the Mod/RM byte
202 /// to specify a destination, which in this case is a register.
206 /// MRMDestMem - This form is used for instructions that use the Mod/RM byte
207 /// to specify a destination, which in this case is memory.
211 /// MRMSrcReg - This form is used for instructions that use the Mod/RM byte
212 /// to specify a source, which in this case is a register.
216 /// MRMSrcMem - This form is used for instructions that use the Mod/RM byte
217 /// to specify a source, which in this case is memory.
221 /// MRM[0-7][rm] - These forms are used to represent instructions that use
222 /// a Mod/RM byte, and use the middle field to hold extended opcode
223 /// information. In the intel manual these are represented as /0, /1, ...
226 // First, instructions that operate on a register r/m operand...
227 MRM0r = 16, MRM1r = 17, MRM2r = 18, MRM3r = 19, // Format /0 /1 /2 /3
228 MRM4r = 20, MRM5r = 21, MRM6r = 22, MRM7r = 23, // Format /4 /5 /6 /7
230 // Next, instructions that operate on a memory r/m operand...
231 MRM0m = 24, MRM1m = 25, MRM2m = 26, MRM3m = 27, // Format /0 /1 /2 /3
232 MRM4m = 28, MRM5m = 29, MRM6m = 30, MRM7m = 31, // Format /4 /5 /6 /7
234 // MRMInitReg - This form is used for instructions whose source and
235 // destinations are the same register.
240 //===------------------------------------------------------------------===//
243 // OpSize - Set if this instruction requires an operand size prefix (0x66),
244 // which most often indicates that the instruction operates on 16 bit data
245 // instead of 32 bit data.
248 // AsSize - Set if this instruction requires an operand size prefix (0x67),
249 // which most often indicates that the instruction address 16 bit address
250 // instead of 32 bit address (or 32 bit address in 64 bit mode).
253 //===------------------------------------------------------------------===//
254 // Op0Mask - There are several prefix bytes that are used to form two byte
255 // opcodes. These are currently 0x0F, 0xF3, and 0xD8-0xDF. This mask is
256 // used to obtain the setting of this field. If no bits in this field is
257 // set, there is no prefix byte for obtaining a multibyte opcode.
260 Op0Mask = 0xF << Op0Shift,
262 // TB - TwoByte - Set if this instruction has a two byte opcode, which
263 // starts with a 0x0F byte before the real opcode.
266 // REP - The 0xF3 prefix byte indicating repetition of the following
270 // D8-DF - These escape opcodes are used by the floating point unit. These
271 // values must remain sequential.
272 D8 = 3 << Op0Shift, D9 = 4 << Op0Shift,
273 DA = 5 << Op0Shift, DB = 6 << Op0Shift,
274 DC = 7 << Op0Shift, DD = 8 << Op0Shift,
275 DE = 9 << Op0Shift, DF = 10 << Op0Shift,
277 // XS, XD - These prefix codes are for single and double precision scalar
278 // floating point operations performed in the SSE registers.
279 XD = 11 << Op0Shift, XS = 12 << Op0Shift,
281 // T8, TA - Prefix after the 0x0F prefix.
282 T8 = 13 << Op0Shift, TA = 14 << Op0Shift,
284 //===------------------------------------------------------------------===//
285 // REX_W - REX prefixes are instruction prefixes used in 64-bit mode.
286 // They are used to specify GPRs and SSE registers, 64-bit operand size,
287 // etc. We only cares about REX.W and REX.R bits and only the former is
288 // statically determined.
291 REX_W = 1 << REXShift,
293 //===------------------------------------------------------------------===//
294 // This three-bit field describes the size of an immediate operand. Zero is
295 // unused so that we can tell if we forgot to set a value.
297 ImmMask = 7 << ImmShift,
298 Imm8 = 1 << ImmShift,
299 Imm16 = 2 << ImmShift,
300 Imm32 = 3 << ImmShift,
301 Imm64 = 4 << ImmShift,
303 //===------------------------------------------------------------------===//
304 // FP Instruction Classification... Zero is non-fp instruction.
306 // FPTypeMask - Mask for all of the FP types...
308 FPTypeMask = 7 << FPTypeShift,
310 // NotFP - The default, set for instructions that do not use FP registers.
311 NotFP = 0 << FPTypeShift,
313 // ZeroArgFP - 0 arg FP instruction which implicitly pushes ST(0), f.e. fld0
314 ZeroArgFP = 1 << FPTypeShift,
316 // OneArgFP - 1 arg FP instructions which implicitly read ST(0), such as fst
317 OneArgFP = 2 << FPTypeShift,
319 // OneArgFPRW - 1 arg FP instruction which implicitly read ST(0) and write a
320 // result back to ST(0). For example, fcos, fsqrt, etc.
322 OneArgFPRW = 3 << FPTypeShift,
324 // TwoArgFP - 2 arg FP instructions which implicitly read ST(0), and an
325 // explicit argument, storing the result to either ST(0) or the implicit
326 // argument. For example: fadd, fsub, fmul, etc...
327 TwoArgFP = 4 << FPTypeShift,
329 // CompareFP - 2 arg FP instructions which implicitly read ST(0) and an
330 // explicit argument, but have no destination. Example: fucom, fucomi, ...
331 CompareFP = 5 << FPTypeShift,
333 // CondMovFP - "2 operand" floating point conditional move instructions.
334 CondMovFP = 6 << FPTypeShift,
336 // SpecialFP - Special instruction forms. Dispatch by opcode explicitly.
337 SpecialFP = 7 << FPTypeShift,
341 LOCK = 1 << LOCKShift,
343 // Segment override prefixes. Currently we just need ability to address
344 // stuff in gs and fs segments.
346 SegOvrMask = 3 << SegOvrShift,
347 FS = 1 << SegOvrShift,
348 GS = 2 << SegOvrShift,
350 // Bits 22 -> 23 are unused
352 OpcodeMask = 0xFF << OpcodeShift
356 const int X86AddrNumOperands = 5;
358 inline static bool isScale(const MachineOperand &MO) {
360 (MO.getImm() == 1 || MO.getImm() == 2 ||
361 MO.getImm() == 4 || MO.getImm() == 8);
364 inline static bool isLeaMem(const MachineInstr *MI, unsigned Op) {
365 if (MI->getOperand(Op).isFI()) return true;
366 return Op+4 <= MI->getNumOperands() &&
367 MI->getOperand(Op ).isReg() && isScale(MI->getOperand(Op+1)) &&
368 MI->getOperand(Op+2).isReg() &&
369 (MI->getOperand(Op+3).isImm() ||
370 MI->getOperand(Op+3).isGlobal() ||
371 MI->getOperand(Op+3).isCPI() ||
372 MI->getOperand(Op+3).isJTI());
375 inline static bool isMem(const MachineInstr *MI, unsigned Op) {
376 if (MI->getOperand(Op).isFI()) return true;
377 return Op+5 <= MI->getNumOperands() &&
378 MI->getOperand(Op+4).isReg() &&
382 class X86InstrInfo : public TargetInstrInfoImpl {
383 X86TargetMachine &TM;
384 const X86RegisterInfo RI;
386 /// RegOp2MemOpTable2Addr, RegOp2MemOpTable0, RegOp2MemOpTable1,
387 /// RegOp2MemOpTable2 - Load / store folding opcode maps.
389 DenseMap<unsigned*, unsigned> RegOp2MemOpTable2Addr;
390 DenseMap<unsigned*, unsigned> RegOp2MemOpTable0;
391 DenseMap<unsigned*, unsigned> RegOp2MemOpTable1;
392 DenseMap<unsigned*, unsigned> RegOp2MemOpTable2;
394 /// MemOp2RegOpTable - Load / store unfolding opcode map.
396 DenseMap<unsigned*, std::pair<unsigned, unsigned> > MemOp2RegOpTable;
399 explicit X86InstrInfo(X86TargetMachine &tm);
401 /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As
402 /// such, whenever a client has an instance of instruction info, it should
403 /// always be able to get register info as well (through this method).
405 virtual const X86RegisterInfo &getRegisterInfo() const { return RI; }
407 /// Return true if the instruction is a register to register move and return
408 /// the source and dest operands and their sub-register indices by reference.
409 virtual bool isMoveInstr(const MachineInstr &MI,
410 unsigned &SrcReg, unsigned &DstReg,
411 unsigned &SrcSubIdx, unsigned &DstSubIdx) const;
413 unsigned isLoadFromStackSlot(const MachineInstr *MI, int &FrameIndex) const;
414 unsigned isStoreToStackSlot(const MachineInstr *MI, int &FrameIndex) const;
416 bool isReallyTriviallyReMaterializable(const MachineInstr *MI) const;
417 void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
418 unsigned DestReg, const MachineInstr *Orig) const;
420 bool isInvariantLoad(const MachineInstr *MI) const;
422 /// convertToThreeAddress - This method must be implemented by targets that
423 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
424 /// may be able to convert a two-address instruction into a true
425 /// three-address instruction on demand. This allows the X86 target (for
426 /// example) to convert ADD and SHL instructions into LEA instructions if they
427 /// would require register copies due to two-addressness.
429 /// This method returns a null pointer if the transformation cannot be
430 /// performed, otherwise it returns the new instruction.
432 virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
433 MachineBasicBlock::iterator &MBBI,
434 LiveVariables *LV) const;
436 /// commuteInstruction - We have a few instructions that must be hacked on to
439 virtual MachineInstr *commuteInstruction(MachineInstr *MI, bool NewMI) const;
442 virtual bool isUnpredicatedTerminator(const MachineInstr* MI) const;
443 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
444 MachineBasicBlock *&FBB,
445 SmallVectorImpl<MachineOperand> &Cond,
446 bool AllowModify) const;
447 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const;
448 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
449 MachineBasicBlock *FBB,
450 const SmallVectorImpl<MachineOperand> &Cond) const;
451 virtual bool copyRegToReg(MachineBasicBlock &MBB,
452 MachineBasicBlock::iterator MI,
453 unsigned DestReg, unsigned SrcReg,
454 const TargetRegisterClass *DestRC,
455 const TargetRegisterClass *SrcRC) const;
456 virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
457 MachineBasicBlock::iterator MI,
458 unsigned SrcReg, bool isKill, int FrameIndex,
459 const TargetRegisterClass *RC) const;
461 virtual void storeRegToAddr(MachineFunction &MF, unsigned SrcReg, bool isKill,
462 SmallVectorImpl<MachineOperand> &Addr,
463 const TargetRegisterClass *RC,
464 SmallVectorImpl<MachineInstr*> &NewMIs) const;
466 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
467 MachineBasicBlock::iterator MI,
468 unsigned DestReg, int FrameIndex,
469 const TargetRegisterClass *RC) const;
471 virtual void loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
472 SmallVectorImpl<MachineOperand> &Addr,
473 const TargetRegisterClass *RC,
474 SmallVectorImpl<MachineInstr*> &NewMIs) const;
476 virtual bool spillCalleeSavedRegisters(MachineBasicBlock &MBB,
477 MachineBasicBlock::iterator MI,
478 const std::vector<CalleeSavedInfo> &CSI) const;
480 virtual bool restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
481 MachineBasicBlock::iterator MI,
482 const std::vector<CalleeSavedInfo> &CSI) const;
484 /// foldMemoryOperand - If this target supports it, fold a load or store of
485 /// the specified stack slot into the specified machine instruction for the
486 /// specified operand(s). If this is possible, the target should perform the
487 /// folding and return true, otherwise it should return false. If it folds
488 /// the instruction, it is likely that the MachineInstruction the iterator
489 /// references has been changed.
490 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
492 const SmallVectorImpl<unsigned> &Ops,
493 int FrameIndex) const;
495 /// foldMemoryOperand - Same as the previous version except it allows folding
496 /// of any load and store from / to any address, not just from a specific
498 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
500 const SmallVectorImpl<unsigned> &Ops,
501 MachineInstr* LoadMI) const;
503 /// canFoldMemoryOperand - Returns true if the specified load / store is
504 /// folding is possible.
505 virtual bool canFoldMemoryOperand(const MachineInstr*,
506 const SmallVectorImpl<unsigned> &) const;
508 /// unfoldMemoryOperand - Separate a single instruction which folded a load or
509 /// a store or a load and a store into two or more instruction. If this is
510 /// possible, returns true as well as the new instructions by reference.
511 virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
512 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
513 SmallVectorImpl<MachineInstr*> &NewMIs) const;
515 virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
516 SmallVectorImpl<SDNode*> &NewNodes) const;
518 /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
519 /// instruction after load / store are unfolded from an instruction of the
520 /// specified opcode. It returns zero if the specified unfolding is not
522 virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
523 bool UnfoldLoad, bool UnfoldStore) const;
525 virtual bool BlockHasNoFallThrough(const MachineBasicBlock &MBB) const;
527 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
529 /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine
530 /// instruction that defines the specified register class.
531 bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const;
533 // getBaseOpcodeFor - This function returns the "base" X86 opcode for the
534 // specified machine instruction.
536 unsigned char getBaseOpcodeFor(const TargetInstrDesc *TID) const {
537 return TID->TSFlags >> X86II::OpcodeShift;
539 unsigned char getBaseOpcodeFor(unsigned Opcode) const {
540 return getBaseOpcodeFor(&get(Opcode));
543 static bool isX86_64NonExtLowByteReg(unsigned reg) {
544 return (reg == X86::SPL || reg == X86::BPL ||
545 reg == X86::SIL || reg == X86::DIL);
548 static unsigned sizeOfImm(const TargetInstrDesc *Desc);
549 static bool isX86_64ExtendedReg(const MachineOperand &MO);
550 static unsigned determineREX(const MachineInstr &MI);
552 /// GetInstSize - Returns the size of the specified MachineInstr.
554 virtual unsigned GetInstSizeInBytes(const MachineInstr *MI) const;
556 /// getGlobalBaseReg - Return a virtual register initialized with the
557 /// the global base register value. Output instructions required to
558 /// initialize the register in the function entry block, if necessary.
560 unsigned getGlobalBaseReg(MachineFunction *MF) const;
563 MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
566 const SmallVectorImpl<MachineOperand> &MOs) const;
569 } // End llvm namespace