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.
72 /// Target Operand Flag enum.
74 //===------------------------------------------------------------------===//
75 // X86 Specific MachineOperand flags.
79 /// MO_GOT_ABSOLUTE_ADDRESS - On a symbol operand, this represents a
81 /// SYMBOL_LABEL + [. - PICBASELABEL]
82 MO_GOT_ABSOLUTE_ADDRESS
= 1,
84 /// MO_PIC_BASE_OFFSET - On a symbol operand this indicates that the
85 /// immediate should get the value of the symbol minus the PIC base label:
86 /// SYMBOL_LABEL - PICBASELABEL
87 MO_PIC_BASE_OFFSET
= 2,
89 /// MO_GOT - On a symbol operand this indicates that the immediate is the
90 /// offset to the GOT entry for the symbol name from the base of the GOT.
92 /// See the X86-64 ELF ABI supplement for more details.
96 /// MO_GOTOFF - On a symbol operand this indicates that the immediate is
97 /// the offset to the location of the symbol name from the base of the GOT.
99 /// See the X86-64 ELF ABI supplement for more details.
100 /// SYMBOL_LABEL @GOTOFF
103 /// MO_GOTPCREL - On a symbol operand this indicates that the immediate is
104 /// offset to the GOT entry for the symbol name from the current code
107 /// See the X86-64 ELF ABI supplement for more details.
108 /// SYMBOL_LABEL @GOTPCREL
111 /// MO_PLT - On a symbol operand this indicates that the immediate is
112 /// offset to the PLT entry of symbol name from the current code location.
114 /// See the X86-64 ELF ABI supplement for more details.
115 /// SYMBOL_LABEL @PLT
118 /// MO_TLSGD - On a symbol operand this indicates that the immediate is
121 /// See 'ELF Handling for Thread-Local Storage' for more details.
122 /// SYMBOL_LABEL @TLSGD
125 /// MO_GOTTPOFF - On a symbol operand this indicates that the immediate is
128 /// See 'ELF Handling for Thread-Local Storage' for more details.
129 /// SYMBOL_LABEL @GOTTPOFF
132 /// MO_INDNTPOFF - On a symbol operand this indicates that the immediate is
135 /// See 'ELF Handling for Thread-Local Storage' for more details.
136 /// SYMBOL_LABEL @INDNTPOFF
139 /// MO_TPOFF - On a symbol operand this indicates that the immediate is
142 /// See 'ELF Handling for Thread-Local Storage' for more details.
143 /// SYMBOL_LABEL @TPOFF
146 /// MO_NTPOFF - On a symbol operand this indicates that the immediate is
149 /// See 'ELF Handling for Thread-Local Storage' for more details.
150 /// SYMBOL_LABEL @NTPOFF
153 /// MO_DLLIMPORT - On a symbol operand "FOO", this indicates that the
154 /// reference is actually to the "__imp_FOO" symbol. This is used for
155 /// dllimport linkage on windows.
158 /// MO_DARWIN_STUB - On a symbol operand "FOO", this indicates that the
159 /// reference is actually to the "FOO$stub" symbol. This is used for calls
160 /// and jumps to external functions on Tiger and before.
163 /// MO_DARWIN_NONLAZY - On a symbol operand "FOO", this indicates that the
164 /// reference is actually to the "FOO$non_lazy_ptr" symbol, which is a
165 /// non-PIC-base-relative reference to a non-hidden dyld lazy pointer stub.
166 MO_DARWIN_NONLAZY
= 14,
168 /// MO_DARWIN_NONLAZY_PIC_BASE - On a symbol operand "FOO", this indicates
169 /// that the reference is actually to "FOO$non_lazy_ptr - PICBASE", which is
170 /// a PIC-base-relative reference to a non-hidden dyld lazy pointer stub.
171 MO_DARWIN_NONLAZY_PIC_BASE
= 15,
173 /// MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE - On a symbol operand "FOO", this
174 /// indicates that the reference is actually to "FOO$non_lazy_ptr -PICBASE",
175 /// which is a PIC-base-relative reference to a hidden dyld lazy pointer
177 MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE
= 16
181 /// isGlobalStubReference - Return true if the specified TargetFlag operand is
182 /// a reference to a stub for a global, not the global itself.
183 inline static bool isGlobalStubReference(unsigned char TargetFlag
) {
184 switch (TargetFlag
) {
185 case X86II::MO_DLLIMPORT
: // dllimport stub.
186 case X86II::MO_GOTPCREL
: // rip-relative GOT reference.
187 case X86II::MO_GOT
: // normal GOT reference.
188 case X86II::MO_DARWIN_NONLAZY_PIC_BASE
: // Normal $non_lazy_ptr ref.
189 case X86II::MO_DARWIN_NONLAZY
: // Normal $non_lazy_ptr ref.
190 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE
: // Hidden $non_lazy_ptr ref.
197 /// isGlobalRelativeToPICBase - Return true if the specified global value
198 /// reference is relative to a 32-bit PIC base (X86ISD::GlobalBaseReg). If this
199 /// is true, the addressing mode has the PIC base register added in (e.g. EBX).
200 inline static bool isGlobalRelativeToPICBase(unsigned char TargetFlag
) {
201 switch (TargetFlag
) {
202 case X86II::MO_GOTOFF
: // isPICStyleGOT: local global.
203 case X86II::MO_GOT
: // isPICStyleGOT: other global.
204 case X86II::MO_PIC_BASE_OFFSET
: // Darwin local global.
205 case X86II::MO_DARWIN_NONLAZY_PIC_BASE
: // Darwin/32 external global.
206 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE
: // Darwin/32 hidden global.
213 /// X86II - This namespace holds all of the target specific flags that
214 /// instruction info tracks.
218 //===------------------------------------------------------------------===//
219 // Instruction encodings. These are the standard/most common forms for X86
223 // PseudoFrm - This represents an instruction that is a pseudo instruction
224 // or one that has not been implemented yet. It is illegal to code generate
225 // it, but tolerated for intermediate implementation stages.
228 /// Raw - This form is for instructions that don't have any operands, so
229 /// they are just a fixed opcode value, like 'leave'.
232 /// AddRegFrm - This form is used for instructions like 'push r32' that have
233 /// their one register operand added to their opcode.
236 /// MRMDestReg - This form is used for instructions that use the Mod/RM byte
237 /// to specify a destination, which in this case is a register.
241 /// MRMDestMem - This form is used for instructions that use the Mod/RM byte
242 /// to specify a destination, which in this case is memory.
246 /// MRMSrcReg - This form is used for instructions that use the Mod/RM byte
247 /// to specify a source, which in this case is a register.
251 /// MRMSrcMem - This form is used for instructions that use the Mod/RM byte
252 /// to specify a source, which in this case is memory.
256 /// MRM[0-7][rm] - These forms are used to represent instructions that use
257 /// a Mod/RM byte, and use the middle field to hold extended opcode
258 /// information. In the intel manual these are represented as /0, /1, ...
261 // First, instructions that operate on a register r/m operand...
262 MRM0r
= 16, MRM1r
= 17, MRM2r
= 18, MRM3r
= 19, // Format /0 /1 /2 /3
263 MRM4r
= 20, MRM5r
= 21, MRM6r
= 22, MRM7r
= 23, // Format /4 /5 /6 /7
265 // Next, instructions that operate on a memory r/m operand...
266 MRM0m
= 24, MRM1m
= 25, MRM2m
= 26, MRM3m
= 27, // Format /0 /1 /2 /3
267 MRM4m
= 28, MRM5m
= 29, MRM6m
= 30, MRM7m
= 31, // Format /4 /5 /6 /7
269 // MRMInitReg - This form is used for instructions whose source and
270 // destinations are the same register.
275 //===------------------------------------------------------------------===//
278 // OpSize - Set if this instruction requires an operand size prefix (0x66),
279 // which most often indicates that the instruction operates on 16 bit data
280 // instead of 32 bit data.
283 // AsSize - Set if this instruction requires an operand size prefix (0x67),
284 // which most often indicates that the instruction address 16 bit address
285 // instead of 32 bit address (or 32 bit address in 64 bit mode).
288 //===------------------------------------------------------------------===//
289 // Op0Mask - There are several prefix bytes that are used to form two byte
290 // opcodes. These are currently 0x0F, 0xF3, and 0xD8-0xDF. This mask is
291 // used to obtain the setting of this field. If no bits in this field is
292 // set, there is no prefix byte for obtaining a multibyte opcode.
295 Op0Mask
= 0xF << Op0Shift
,
297 // TB - TwoByte - Set if this instruction has a two byte opcode, which
298 // starts with a 0x0F byte before the real opcode.
301 // REP - The 0xF3 prefix byte indicating repetition of the following
305 // D8-DF - These escape opcodes are used by the floating point unit. These
306 // values must remain sequential.
307 D8
= 3 << Op0Shift
, D9
= 4 << Op0Shift
,
308 DA
= 5 << Op0Shift
, DB
= 6 << Op0Shift
,
309 DC
= 7 << Op0Shift
, DD
= 8 << Op0Shift
,
310 DE
= 9 << Op0Shift
, DF
= 10 << Op0Shift
,
312 // XS, XD - These prefix codes are for single and double precision scalar
313 // floating point operations performed in the SSE registers.
314 XD
= 11 << Op0Shift
, XS
= 12 << Op0Shift
,
316 // T8, TA - Prefix after the 0x0F prefix.
317 T8
= 13 << Op0Shift
, TA
= 14 << Op0Shift
,
319 // TF - Prefix before and after 0x0F
322 //===------------------------------------------------------------------===//
323 // REX_W - REX prefixes are instruction prefixes used in 64-bit mode.
324 // They are used to specify GPRs and SSE registers, 64-bit operand size,
325 // etc. We only cares about REX.W and REX.R bits and only the former is
326 // statically determined.
329 REX_W
= 1 << REXShift
,
331 //===------------------------------------------------------------------===//
332 // This three-bit field describes the size of an immediate operand. Zero is
333 // unused so that we can tell if we forgot to set a value.
335 ImmMask
= 7 << ImmShift
,
336 Imm8
= 1 << ImmShift
,
337 Imm16
= 2 << ImmShift
,
338 Imm32
= 3 << ImmShift
,
339 Imm64
= 4 << ImmShift
,
341 //===------------------------------------------------------------------===//
342 // FP Instruction Classification... Zero is non-fp instruction.
344 // FPTypeMask - Mask for all of the FP types...
346 FPTypeMask
= 7 << FPTypeShift
,
348 // NotFP - The default, set for instructions that do not use FP registers.
349 NotFP
= 0 << FPTypeShift
,
351 // ZeroArgFP - 0 arg FP instruction which implicitly pushes ST(0), f.e. fld0
352 ZeroArgFP
= 1 << FPTypeShift
,
354 // OneArgFP - 1 arg FP instructions which implicitly read ST(0), such as fst
355 OneArgFP
= 2 << FPTypeShift
,
357 // OneArgFPRW - 1 arg FP instruction which implicitly read ST(0) and write a
358 // result back to ST(0). For example, fcos, fsqrt, etc.
360 OneArgFPRW
= 3 << FPTypeShift
,
362 // TwoArgFP - 2 arg FP instructions which implicitly read ST(0), and an
363 // explicit argument, storing the result to either ST(0) or the implicit
364 // argument. For example: fadd, fsub, fmul, etc...
365 TwoArgFP
= 4 << FPTypeShift
,
367 // CompareFP - 2 arg FP instructions which implicitly read ST(0) and an
368 // explicit argument, but have no destination. Example: fucom, fucomi, ...
369 CompareFP
= 5 << FPTypeShift
,
371 // CondMovFP - "2 operand" floating point conditional move instructions.
372 CondMovFP
= 6 << FPTypeShift
,
374 // SpecialFP - Special instruction forms. Dispatch by opcode explicitly.
375 SpecialFP
= 7 << FPTypeShift
,
379 LOCK
= 1 << LOCKShift
,
381 // Segment override prefixes. Currently we just need ability to address
382 // stuff in gs and fs segments.
384 SegOvrMask
= 3 << SegOvrShift
,
385 FS
= 1 << SegOvrShift
,
386 GS
= 2 << SegOvrShift
,
388 // Bits 22 -> 23 are unused
390 OpcodeMask
= 0xFF << OpcodeShift
394 const int X86AddrNumOperands
= 5;
396 inline static bool isScale(const MachineOperand
&MO
) {
398 (MO
.getImm() == 1 || MO
.getImm() == 2 ||
399 MO
.getImm() == 4 || MO
.getImm() == 8);
402 inline static bool isLeaMem(const MachineInstr
*MI
, unsigned Op
) {
403 if (MI
->getOperand(Op
).isFI()) return true;
404 return Op
+4 <= MI
->getNumOperands() &&
405 MI
->getOperand(Op
).isReg() && isScale(MI
->getOperand(Op
+1)) &&
406 MI
->getOperand(Op
+2).isReg() &&
407 (MI
->getOperand(Op
+3).isImm() ||
408 MI
->getOperand(Op
+3).isGlobal() ||
409 MI
->getOperand(Op
+3).isCPI() ||
410 MI
->getOperand(Op
+3).isJTI());
413 inline static bool isMem(const MachineInstr
*MI
, unsigned Op
) {
414 if (MI
->getOperand(Op
).isFI()) return true;
415 return Op
+5 <= MI
->getNumOperands() &&
416 MI
->getOperand(Op
+4).isReg() &&
420 class X86InstrInfo
: public TargetInstrInfoImpl
{
421 X86TargetMachine
&TM
;
422 const X86RegisterInfo RI
;
424 /// RegOp2MemOpTable2Addr, RegOp2MemOpTable0, RegOp2MemOpTable1,
425 /// RegOp2MemOpTable2 - Load / store folding opcode maps.
427 DenseMap
<unsigned*, std::pair
<unsigned,unsigned> > RegOp2MemOpTable2Addr
;
428 DenseMap
<unsigned*, std::pair
<unsigned,unsigned> > RegOp2MemOpTable0
;
429 DenseMap
<unsigned*, std::pair
<unsigned,unsigned> > RegOp2MemOpTable1
;
430 DenseMap
<unsigned*, std::pair
<unsigned,unsigned> > RegOp2MemOpTable2
;
432 /// MemOp2RegOpTable - Load / store unfolding opcode map.
434 DenseMap
<unsigned*, std::pair
<unsigned, unsigned> > MemOp2RegOpTable
;
437 explicit X86InstrInfo(X86TargetMachine
&tm
);
439 /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As
440 /// such, whenever a client has an instance of instruction info, it should
441 /// always be able to get register info as well (through this method).
443 virtual const X86RegisterInfo
&getRegisterInfo() const { return RI
; }
445 /// Return true if the instruction is a register to register move and return
446 /// the source and dest operands and their sub-register indices by reference.
447 virtual bool isMoveInstr(const MachineInstr
&MI
,
448 unsigned &SrcReg
, unsigned &DstReg
,
449 unsigned &SrcSubIdx
, unsigned &DstSubIdx
) const;
451 unsigned isLoadFromStackSlot(const MachineInstr
*MI
, int &FrameIndex
) const;
452 unsigned isStoreToStackSlot(const MachineInstr
*MI
, int &FrameIndex
) const;
454 bool isReallyTriviallyReMaterializable(const MachineInstr
*MI
) const;
455 void reMaterialize(MachineBasicBlock
&MBB
, MachineBasicBlock::iterator MI
,
456 unsigned DestReg
, unsigned SubIdx
,
457 const MachineInstr
*Orig
) const;
459 bool isInvariantLoad(const MachineInstr
*MI
) const;
461 /// convertToThreeAddress - This method must be implemented by targets that
462 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
463 /// may be able to convert a two-address instruction into a true
464 /// three-address instruction on demand. This allows the X86 target (for
465 /// example) to convert ADD and SHL instructions into LEA instructions if they
466 /// would require register copies due to two-addressness.
468 /// This method returns a null pointer if the transformation cannot be
469 /// performed, otherwise it returns the new instruction.
471 virtual MachineInstr
*convertToThreeAddress(MachineFunction::iterator
&MFI
,
472 MachineBasicBlock::iterator
&MBBI
,
473 LiveVariables
*LV
) const;
475 /// commuteInstruction - We have a few instructions that must be hacked on to
478 virtual MachineInstr
*commuteInstruction(MachineInstr
*MI
, bool NewMI
) const;
481 virtual bool isUnpredicatedTerminator(const MachineInstr
* MI
) const;
482 virtual bool AnalyzeBranch(MachineBasicBlock
&MBB
, MachineBasicBlock
*&TBB
,
483 MachineBasicBlock
*&FBB
,
484 SmallVectorImpl
<MachineOperand
> &Cond
,
485 bool AllowModify
) const;
486 virtual unsigned RemoveBranch(MachineBasicBlock
&MBB
) const;
487 virtual unsigned InsertBranch(MachineBasicBlock
&MBB
, MachineBasicBlock
*TBB
,
488 MachineBasicBlock
*FBB
,
489 const SmallVectorImpl
<MachineOperand
> &Cond
) const;
490 virtual bool copyRegToReg(MachineBasicBlock
&MBB
,
491 MachineBasicBlock::iterator MI
,
492 unsigned DestReg
, unsigned SrcReg
,
493 const TargetRegisterClass
*DestRC
,
494 const TargetRegisterClass
*SrcRC
) const;
495 virtual void storeRegToStackSlot(MachineBasicBlock
&MBB
,
496 MachineBasicBlock::iterator MI
,
497 unsigned SrcReg
, bool isKill
, int FrameIndex
,
498 const TargetRegisterClass
*RC
) const;
500 virtual void storeRegToAddr(MachineFunction
&MF
, unsigned SrcReg
, bool isKill
,
501 SmallVectorImpl
<MachineOperand
> &Addr
,
502 const TargetRegisterClass
*RC
,
503 SmallVectorImpl
<MachineInstr
*> &NewMIs
) const;
505 virtual void loadRegFromStackSlot(MachineBasicBlock
&MBB
,
506 MachineBasicBlock::iterator MI
,
507 unsigned DestReg
, int FrameIndex
,
508 const TargetRegisterClass
*RC
) const;
510 virtual void loadRegFromAddr(MachineFunction
&MF
, unsigned DestReg
,
511 SmallVectorImpl
<MachineOperand
> &Addr
,
512 const TargetRegisterClass
*RC
,
513 SmallVectorImpl
<MachineInstr
*> &NewMIs
) const;
515 virtual bool spillCalleeSavedRegisters(MachineBasicBlock
&MBB
,
516 MachineBasicBlock::iterator MI
,
517 const std::vector
<CalleeSavedInfo
> &CSI
) const;
519 virtual bool restoreCalleeSavedRegisters(MachineBasicBlock
&MBB
,
520 MachineBasicBlock::iterator MI
,
521 const std::vector
<CalleeSavedInfo
> &CSI
) const;
523 /// foldMemoryOperand - If this target supports it, fold a load or store of
524 /// the specified stack slot into the specified machine instruction for the
525 /// specified operand(s). If this is possible, the target should perform the
526 /// folding and return true, otherwise it should return false. If it folds
527 /// the instruction, it is likely that the MachineInstruction the iterator
528 /// references has been changed.
529 virtual MachineInstr
* foldMemoryOperandImpl(MachineFunction
&MF
,
531 const SmallVectorImpl
<unsigned> &Ops
,
532 int FrameIndex
) const;
534 /// foldMemoryOperand - Same as the previous version except it allows folding
535 /// of any load and store from / to any address, not just from a specific
537 virtual MachineInstr
* foldMemoryOperandImpl(MachineFunction
&MF
,
539 const SmallVectorImpl
<unsigned> &Ops
,
540 MachineInstr
* LoadMI
) const;
542 /// canFoldMemoryOperand - Returns true if the specified load / store is
543 /// folding is possible.
544 virtual bool canFoldMemoryOperand(const MachineInstr
*,
545 const SmallVectorImpl
<unsigned> &) const;
547 /// unfoldMemoryOperand - Separate a single instruction which folded a load or
548 /// a store or a load and a store into two or more instruction. If this is
549 /// possible, returns true as well as the new instructions by reference.
550 virtual bool unfoldMemoryOperand(MachineFunction
&MF
, MachineInstr
*MI
,
551 unsigned Reg
, bool UnfoldLoad
, bool UnfoldStore
,
552 SmallVectorImpl
<MachineInstr
*> &NewMIs
) const;
554 virtual bool unfoldMemoryOperand(SelectionDAG
&DAG
, SDNode
*N
,
555 SmallVectorImpl
<SDNode
*> &NewNodes
) const;
557 /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
558 /// instruction after load / store are unfolded from an instruction of the
559 /// specified opcode. It returns zero if the specified unfolding is not
561 virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc
,
562 bool UnfoldLoad
, bool UnfoldStore
) const;
564 virtual bool BlockHasNoFallThrough(const MachineBasicBlock
&MBB
) const;
566 bool ReverseBranchCondition(SmallVectorImpl
<MachineOperand
> &Cond
) const;
568 /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine
569 /// instruction that defines the specified register class.
570 bool isSafeToMoveRegClassDefs(const TargetRegisterClass
*RC
) const;
572 // getBaseOpcodeFor - This function returns the "base" X86 opcode for the
573 // specified machine instruction.
575 unsigned char getBaseOpcodeFor(const TargetInstrDesc
*TID
) const {
576 return TID
->TSFlags
>> X86II::OpcodeShift
;
578 unsigned char getBaseOpcodeFor(unsigned Opcode
) const {
579 return getBaseOpcodeFor(&get(Opcode
));
582 static bool isX86_64NonExtLowByteReg(unsigned reg
) {
583 return (reg
== X86::SPL
|| reg
== X86::BPL
||
584 reg
== X86::SIL
|| reg
== X86::DIL
);
587 static unsigned sizeOfImm(const TargetInstrDesc
*Desc
);
588 static bool isX86_64ExtendedReg(const MachineOperand
&MO
);
589 static unsigned determineREX(const MachineInstr
&MI
);
591 /// GetInstSize - Returns the size of the specified MachineInstr.
593 virtual unsigned GetInstSizeInBytes(const MachineInstr
*MI
) const;
595 /// getGlobalBaseReg - Return a virtual register initialized with the
596 /// the global base register value. Output instructions required to
597 /// initialize the register in the function entry block, if necessary.
599 unsigned getGlobalBaseReg(MachineFunction
*MF
) const;
602 MachineInstr
* foldMemoryOperandImpl(MachineFunction
&MF
,
605 const SmallVectorImpl
<MachineOperand
> &MOs
,
606 unsigned Size
, unsigned Alignment
) const;
609 } // End llvm namespace