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"
23 class X86RegisterInfo;
24 class X86TargetMachine;
27 // Enums for memory operand decoding. Each memory operand is represented with
28 // a 5 operand sequence in the form:
29 // [BaseReg, ScaleAmt, IndexReg, Disp, Segment]
30 // These enums help decode this.
37 /// AddrSegmentReg - The operand # of the segment in the memory operand.
40 /// AddrNumOperands - Total number of operands in a memory reference.
45 // X86 specific condition code. These correspond to X86_*_COND in
46 // X86InstrInfo.td. They must be kept in synch.
65 // Artificial condition codes. These are used by AnalyzeBranch
66 // to indicate a block terminated with two conditional branches to
67 // the same location. This occurs in code using FCMP_OEQ or FCMP_UNE,
68 // which can't be represented on x86 with a single condition. These
69 // are never used in MachineInstrs.
76 // Turn condition code into conditional branch opcode.
77 unsigned GetCondBranchFromCond(CondCode CC);
79 /// GetOppositeBranchCondition - Return the inverse of the specified cond,
80 /// e.g. turning COND_E to COND_NE.
81 CondCode GetOppositeBranchCondition(X86::CondCode CC);
85 /// X86II - This namespace holds all of the target specific flags that
86 /// instruction info tracks.
89 /// Target Operand Flag enum.
91 //===------------------------------------------------------------------===//
92 // X86 Specific MachineOperand flags.
96 /// MO_GOT_ABSOLUTE_ADDRESS - On a symbol operand, this represents a
98 /// SYMBOL_LABEL + [. - PICBASELABEL]
99 MO_GOT_ABSOLUTE_ADDRESS,
101 /// MO_PIC_BASE_OFFSET - On a symbol operand this indicates that the
102 /// immediate should get the value of the symbol minus the PIC base label:
103 /// SYMBOL_LABEL - PICBASELABEL
106 /// MO_GOT - On a symbol operand this indicates that the immediate is the
107 /// offset to the GOT entry for the symbol name from the base of the GOT.
109 /// See the X86-64 ELF ABI supplement for more details.
110 /// SYMBOL_LABEL @GOT
113 /// MO_GOTOFF - On a symbol operand this indicates that the immediate is
114 /// the offset to the location of the symbol name from the base of the GOT.
116 /// See the X86-64 ELF ABI supplement for more details.
117 /// SYMBOL_LABEL @GOTOFF
120 /// MO_GOTPCREL - On a symbol operand this indicates that the immediate is
121 /// offset to the GOT entry for the symbol name from the current code
124 /// See the X86-64 ELF ABI supplement for more details.
125 /// SYMBOL_LABEL @GOTPCREL
128 /// MO_PLT - On a symbol operand this indicates that the immediate is
129 /// offset to the PLT entry of symbol name from the current code location.
131 /// See the X86-64 ELF ABI supplement for more details.
132 /// SYMBOL_LABEL @PLT
135 /// MO_TLSGD - On a symbol operand this indicates that the immediate is
138 /// See 'ELF Handling for Thread-Local Storage' for more details.
139 /// SYMBOL_LABEL @TLSGD
142 /// MO_GOTTPOFF - On a symbol operand this indicates that the immediate is
145 /// See 'ELF Handling for Thread-Local Storage' for more details.
146 /// SYMBOL_LABEL @GOTTPOFF
149 /// MO_INDNTPOFF - On a symbol operand this indicates that the immediate is
152 /// See 'ELF Handling for Thread-Local Storage' for more details.
153 /// SYMBOL_LABEL @INDNTPOFF
156 /// MO_TPOFF - On a symbol operand this indicates that the immediate is
159 /// See 'ELF Handling for Thread-Local Storage' for more details.
160 /// SYMBOL_LABEL @TPOFF
163 /// MO_NTPOFF - On a symbol operand this indicates that the immediate is
166 /// See 'ELF Handling for Thread-Local Storage' for more details.
167 /// SYMBOL_LABEL @NTPOFF
170 /// MO_DLLIMPORT - On a symbol operand "FOO", this indicates that the
171 /// reference is actually to the "__imp_FOO" symbol. This is used for
172 /// dllimport linkage on windows.
175 /// MO_DARWIN_STUB - On a symbol operand "FOO", this indicates that the
176 /// reference is actually to the "FOO$stub" symbol. This is used for calls
177 /// and jumps to external functions on Tiger and before.
180 /// MO_DARWIN_NONLAZY - On a symbol operand "FOO", this indicates that the
181 /// reference is actually to the "FOO$non_lazy_ptr" symbol, which is a
182 /// non-PIC-base-relative reference to a non-hidden dyld lazy pointer stub.
185 /// MO_DARWIN_NONLAZY_PIC_BASE - On a symbol operand "FOO", this indicates
186 /// that the reference is actually to "FOO$non_lazy_ptr - PICBASE", which is
187 /// a PIC-base-relative reference to a non-hidden dyld lazy pointer stub.
188 MO_DARWIN_NONLAZY_PIC_BASE,
190 /// MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE - On a symbol operand "FOO", this
191 /// indicates that the reference is actually to "FOO$non_lazy_ptr -PICBASE",
192 /// which is a PIC-base-relative reference to a hidden dyld lazy pointer
194 MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE,
196 /// MO_TLVP - On a symbol operand this indicates that the immediate is
199 /// This is the TLS offset for the Darwin TLS mechanism.
202 /// MO_TLVP_PIC_BASE - On a symbol operand this indicates that the immediate
203 /// is some TLS offset from the picbase.
205 /// This is the 32-bit TLS offset for Darwin TLS in PIC mode.
210 /// isGlobalStubReference - Return true if the specified TargetFlag operand is
211 /// a reference to a stub for a global, not the global itself.
212 inline static bool isGlobalStubReference(unsigned char TargetFlag) {
213 switch (TargetFlag) {
214 case X86II::MO_DLLIMPORT: // dllimport stub.
215 case X86II::MO_GOTPCREL: // rip-relative GOT reference.
216 case X86II::MO_GOT: // normal GOT reference.
217 case X86II::MO_DARWIN_NONLAZY_PIC_BASE: // Normal $non_lazy_ptr ref.
218 case X86II::MO_DARWIN_NONLAZY: // Normal $non_lazy_ptr ref.
219 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Hidden $non_lazy_ptr ref.
226 /// isGlobalRelativeToPICBase - Return true if the specified global value
227 /// reference is relative to a 32-bit PIC base (X86ISD::GlobalBaseReg). If this
228 /// is true, the addressing mode has the PIC base register added in (e.g. EBX).
229 inline static bool isGlobalRelativeToPICBase(unsigned char TargetFlag) {
230 switch (TargetFlag) {
231 case X86II::MO_GOTOFF: // isPICStyleGOT: local global.
232 case X86II::MO_GOT: // isPICStyleGOT: other global.
233 case X86II::MO_PIC_BASE_OFFSET: // Darwin local global.
234 case X86II::MO_DARWIN_NONLAZY_PIC_BASE: // Darwin/32 external global.
235 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Darwin/32 hidden global.
236 case X86II::MO_TLVP: // ??? Pretty sure..
243 /// X86II - This namespace holds all of the target specific flags that
244 /// instruction info tracks.
248 //===------------------------------------------------------------------===//
249 // Instruction encodings. These are the standard/most common forms for X86
253 // PseudoFrm - This represents an instruction that is a pseudo instruction
254 // or one that has not been implemented yet. It is illegal to code generate
255 // it, but tolerated for intermediate implementation stages.
258 /// Raw - This form is for instructions that don't have any operands, so
259 /// they are just a fixed opcode value, like 'leave'.
262 /// AddRegFrm - This form is used for instructions like 'push r32' that have
263 /// their one register operand added to their opcode.
266 /// MRMDestReg - This form is used for instructions that use the Mod/RM byte
267 /// to specify a destination, which in this case is a register.
271 /// MRMDestMem - This form is used for instructions that use the Mod/RM byte
272 /// to specify a destination, which in this case is memory.
276 /// MRMSrcReg - This form is used for instructions that use the Mod/RM byte
277 /// to specify a source, which in this case is a register.
281 /// MRMSrcMem - This form is used for instructions that use the Mod/RM byte
282 /// to specify a source, which in this case is memory.
286 /// MRM[0-7][rm] - These forms are used to represent instructions that use
287 /// a Mod/RM byte, and use the middle field to hold extended opcode
288 /// information. In the intel manual these are represented as /0, /1, ...
291 // First, instructions that operate on a register r/m operand...
292 MRM0r = 16, MRM1r = 17, MRM2r = 18, MRM3r = 19, // Format /0 /1 /2 /3
293 MRM4r = 20, MRM5r = 21, MRM6r = 22, MRM7r = 23, // Format /4 /5 /6 /7
295 // Next, instructions that operate on a memory r/m operand...
296 MRM0m = 24, MRM1m = 25, MRM2m = 26, MRM3m = 27, // Format /0 /1 /2 /3
297 MRM4m = 28, MRM5m = 29, MRM6m = 30, MRM7m = 31, // Format /4 /5 /6 /7
299 // MRMInitReg - This form is used for instructions whose source and
300 // destinations are the same register.
303 //// MRM_C1 - A mod/rm byte of exactly 0xC1.
317 //===------------------------------------------------------------------===//
320 // OpSize - Set if this instruction requires an operand size prefix (0x66),
321 // which most often indicates that the instruction operates on 16 bit data
322 // instead of 32 bit data.
325 // AsSize - Set if this instruction requires an operand size prefix (0x67),
326 // which most often indicates that the instruction address 16 bit address
327 // instead of 32 bit address (or 32 bit address in 64 bit mode).
330 //===------------------------------------------------------------------===//
331 // Op0Mask - There are several prefix bytes that are used to form two byte
332 // opcodes. These are currently 0x0F, 0xF3, and 0xD8-0xDF. This mask is
333 // used to obtain the setting of this field. If no bits in this field is
334 // set, there is no prefix byte for obtaining a multibyte opcode.
337 Op0Mask = 0xF << Op0Shift,
339 // TB - TwoByte - Set if this instruction has a two byte opcode, which
340 // starts with a 0x0F byte before the real opcode.
343 // REP - The 0xF3 prefix byte indicating repetition of the following
347 // D8-DF - These escape opcodes are used by the floating point unit. These
348 // values must remain sequential.
349 D8 = 3 << Op0Shift, D9 = 4 << Op0Shift,
350 DA = 5 << Op0Shift, DB = 6 << Op0Shift,
351 DC = 7 << Op0Shift, DD = 8 << Op0Shift,
352 DE = 9 << Op0Shift, DF = 10 << Op0Shift,
354 // XS, XD - These prefix codes are for single and double precision scalar
355 // floating point operations performed in the SSE registers.
356 XD = 11 << Op0Shift, XS = 12 << Op0Shift,
358 // T8, TA - Prefix after the 0x0F prefix.
359 T8 = 13 << Op0Shift, TA = 14 << Op0Shift,
361 // TF - Prefix before and after 0x0F
364 //===------------------------------------------------------------------===//
365 // REX_W - REX prefixes are instruction prefixes used in 64-bit mode.
366 // They are used to specify GPRs and SSE registers, 64-bit operand size,
367 // etc. We only cares about REX.W and REX.R bits and only the former is
368 // statically determined.
371 REX_W = 1 << REXShift,
373 //===------------------------------------------------------------------===//
374 // This three-bit field describes the size of an immediate operand. Zero is
375 // unused so that we can tell if we forgot to set a value.
377 ImmMask = 7 << ImmShift,
378 Imm8 = 1 << ImmShift,
379 Imm8PCRel = 2 << ImmShift,
380 Imm16 = 3 << ImmShift,
381 Imm16PCRel = 4 << ImmShift,
382 Imm32 = 5 << ImmShift,
383 Imm32PCRel = 6 << ImmShift,
384 Imm64 = 7 << ImmShift,
386 //===------------------------------------------------------------------===//
387 // FP Instruction Classification... Zero is non-fp instruction.
389 // FPTypeMask - Mask for all of the FP types...
391 FPTypeMask = 7 << FPTypeShift,
393 // NotFP - The default, set for instructions that do not use FP registers.
394 NotFP = 0 << FPTypeShift,
396 // ZeroArgFP - 0 arg FP instruction which implicitly pushes ST(0), f.e. fld0
397 ZeroArgFP = 1 << FPTypeShift,
399 // OneArgFP - 1 arg FP instructions which implicitly read ST(0), such as fst
400 OneArgFP = 2 << FPTypeShift,
402 // OneArgFPRW - 1 arg FP instruction which implicitly read ST(0) and write a
403 // result back to ST(0). For example, fcos, fsqrt, etc.
405 OneArgFPRW = 3 << FPTypeShift,
407 // TwoArgFP - 2 arg FP instructions which implicitly read ST(0), and an
408 // explicit argument, storing the result to either ST(0) or the implicit
409 // argument. For example: fadd, fsub, fmul, etc...
410 TwoArgFP = 4 << FPTypeShift,
412 // CompareFP - 2 arg FP instructions which implicitly read ST(0) and an
413 // explicit argument, but have no destination. Example: fucom, fucomi, ...
414 CompareFP = 5 << FPTypeShift,
416 // CondMovFP - "2 operand" floating point conditional move instructions.
417 CondMovFP = 6 << FPTypeShift,
419 // SpecialFP - Special instruction forms. Dispatch by opcode explicitly.
420 SpecialFP = 7 << FPTypeShift,
424 LOCK = 1 << LOCKShift,
426 // Segment override prefixes. Currently we just need ability to address
427 // stuff in gs and fs segments.
429 SegOvrMask = 3 << SegOvrShift,
430 FS = 1 << SegOvrShift,
431 GS = 2 << SegOvrShift,
433 // Execution domain for SSE instructions in bits 22, 23.
434 // 0 in bits 22-23 means normal, non-SSE instruction.
438 OpcodeMask = 0xFF << OpcodeShift,
440 //===------------------------------------------------------------------===//
441 // VEX - The opcode prefix used by AVX instructions
444 // VEX_W - Has a opcode specific functionality, but is used in the same
445 // way as REX_W is for regular SSE instructions.
448 // VEX_4V - Used to specify an additional AVX/SSE register. Several 2
449 // address instructions in SSE are represented as 3 address ones in AVX
450 // and the additional register is encoded in VEX_VVVV prefix.
453 // VEX_I8IMM - Specifies that the last register used in a AVX instruction,
454 // must be encoded in the i8 immediate field. This usually happens in
455 // instructions with 4 operands.
456 VEX_I8IMM = 1ULL << 35,
458 // VEX_L - Stands for a bit in the VEX opcode prefix meaning the current
459 // instruction uses 256-bit wide registers. This is usually auto detected if
460 // a VR256 register is used, but some AVX instructions also have this field
461 // marked when using a f256 memory references.
465 // getBaseOpcodeFor - This function returns the "base" X86 opcode for the
466 // specified machine instruction.
468 static inline unsigned char getBaseOpcodeFor(uint64_t TSFlags) {
469 return TSFlags >> X86II::OpcodeShift;
472 static inline bool hasImm(uint64_t TSFlags) {
473 return (TSFlags & X86II::ImmMask) != 0;
476 /// getSizeOfImm - Decode the "size of immediate" field from the TSFlags field
477 /// of the specified instruction.
478 static inline unsigned getSizeOfImm(uint64_t TSFlags) {
479 switch (TSFlags & X86II::ImmMask) {
480 default: assert(0 && "Unknown immediate size");
482 case X86II::Imm8PCRel: return 1;
484 case X86II::Imm16PCRel: return 2;
486 case X86II::Imm32PCRel: return 4;
487 case X86II::Imm64: return 8;
491 /// isImmPCRel - Return true if the immediate of the specified instruction's
492 /// TSFlags indicates that it is pc relative.
493 static inline unsigned isImmPCRel(uint64_t TSFlags) {
494 switch (TSFlags & X86II::ImmMask) {
495 default: assert(0 && "Unknown immediate size");
496 case X86II::Imm8PCRel:
497 case X86II::Imm16PCRel:
498 case X86II::Imm32PCRel:
508 /// getMemoryOperandNo - The function returns the MCInst operand # for the
509 /// first field of the memory operand. If the instruction doesn't have a
510 /// memory operand, this returns -1.
512 /// Note that this ignores tied operands. If there is a tied register which
513 /// is duplicated in the MCInst (e.g. "EAX = addl EAX, [mem]") it is only
514 /// counted as one operand.
516 static inline int getMemoryOperandNo(uint64_t TSFlags) {
517 switch (TSFlags & X86II::FormMask) {
518 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this form");
519 default: assert(0 && "Unknown FormMask value in getMemoryOperandNo!");
522 case X86II::AddRegFrm:
523 case X86II::MRMDestReg:
524 case X86II::MRMSrcReg:
526 case X86II::MRMDestMem:
528 case X86II::MRMSrcMem: {
529 bool HasVEX_4V = TSFlags & X86II::VEX_4V;
530 unsigned FirstMemOp = 1;
532 ++FirstMemOp;// Skip the register source (which is encoded in VEX_VVVV).
534 // FIXME: Maybe lea should have its own form? This is a horrible hack.
535 //if (Opcode == X86::LEA64r || Opcode == X86::LEA64_32r ||
536 // Opcode == X86::LEA16r || Opcode == X86::LEA32r)
539 case X86II::MRM0r: case X86II::MRM1r:
540 case X86II::MRM2r: case X86II::MRM3r:
541 case X86II::MRM4r: case X86II::MRM5r:
542 case X86II::MRM6r: case X86II::MRM7r:
544 case X86II::MRM0m: case X86II::MRM1m:
545 case X86II::MRM2m: case X86II::MRM3m:
546 case X86II::MRM4m: case X86II::MRM5m:
547 case X86II::MRM6m: case X86II::MRM7m:
564 inline static bool isScale(const MachineOperand &MO) {
566 (MO.getImm() == 1 || MO.getImm() == 2 ||
567 MO.getImm() == 4 || MO.getImm() == 8);
570 inline static bool isLeaMem(const MachineInstr *MI, unsigned Op) {
571 if (MI->getOperand(Op).isFI()) return true;
572 return Op+4 <= MI->getNumOperands() &&
573 MI->getOperand(Op ).isReg() && isScale(MI->getOperand(Op+1)) &&
574 MI->getOperand(Op+2).isReg() &&
575 (MI->getOperand(Op+3).isImm() ||
576 MI->getOperand(Op+3).isGlobal() ||
577 MI->getOperand(Op+3).isCPI() ||
578 MI->getOperand(Op+3).isJTI());
581 inline static bool isMem(const MachineInstr *MI, unsigned Op) {
582 if (MI->getOperand(Op).isFI()) return true;
583 return Op+5 <= MI->getNumOperands() &&
584 MI->getOperand(Op+4).isReg() &&
588 class X86InstrInfo : public TargetInstrInfoImpl {
589 X86TargetMachine &TM;
590 const X86RegisterInfo RI;
592 /// RegOp2MemOpTable2Addr, RegOp2MemOpTable0, RegOp2MemOpTable1,
593 /// RegOp2MemOpTable2 - Load / store folding opcode maps.
595 DenseMap<unsigned*, std::pair<unsigned,unsigned> > RegOp2MemOpTable2Addr;
596 DenseMap<unsigned*, std::pair<unsigned,unsigned> > RegOp2MemOpTable0;
597 DenseMap<unsigned*, std::pair<unsigned,unsigned> > RegOp2MemOpTable1;
598 DenseMap<unsigned*, std::pair<unsigned,unsigned> > RegOp2MemOpTable2;
600 /// MemOp2RegOpTable - Load / store unfolding opcode map.
602 DenseMap<unsigned*, std::pair<unsigned, unsigned> > MemOp2RegOpTable;
605 explicit X86InstrInfo(X86TargetMachine &tm);
607 /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As
608 /// such, whenever a client has an instance of instruction info, it should
609 /// always be able to get register info as well (through this method).
611 virtual const X86RegisterInfo &getRegisterInfo() const { return RI; }
613 /// isCoalescableExtInstr - Return true if the instruction is a "coalescable"
614 /// extension instruction. That is, it's like a copy where it's legal for the
615 /// source to overlap the destination. e.g. X86::MOVSX64rr32. If this returns
616 /// true, then it's expected the pre-extension value is available as a subreg
617 /// of the result register. This also returns the sub-register index in
619 virtual bool isCoalescableExtInstr(const MachineInstr &MI,
620 unsigned &SrcReg, unsigned &DstReg,
621 unsigned &SubIdx) const;
623 unsigned isLoadFromStackSlot(const MachineInstr *MI, int &FrameIndex) const;
624 /// isLoadFromStackSlotPostFE - Check for post-frame ptr elimination
625 /// stack locations as well. This uses a heuristic so it isn't
626 /// reliable for correctness.
627 unsigned isLoadFromStackSlotPostFE(const MachineInstr *MI,
628 int &FrameIndex) const;
630 /// hasLoadFromStackSlot - If the specified machine instruction has
631 /// a load from a stack slot, return true along with the FrameIndex
632 /// of the loaded stack slot and the machine mem operand containing
633 /// the reference. If not, return false. Unlike
634 /// isLoadFromStackSlot, this returns true for any instructions that
635 /// loads from the stack. This is a hint only and may not catch all
637 bool hasLoadFromStackSlot(const MachineInstr *MI,
638 const MachineMemOperand *&MMO,
639 int &FrameIndex) const;
641 unsigned isStoreToStackSlot(const MachineInstr *MI, int &FrameIndex) const;
642 /// isStoreToStackSlotPostFE - Check for post-frame ptr elimination
643 /// stack locations as well. This uses a heuristic so it isn't
644 /// reliable for correctness.
645 unsigned isStoreToStackSlotPostFE(const MachineInstr *MI,
646 int &FrameIndex) const;
648 /// hasStoreToStackSlot - If the specified machine instruction has a
649 /// store to a stack slot, return true along with the FrameIndex of
650 /// the loaded stack slot and the machine mem operand containing the
651 /// reference. If not, return false. Unlike isStoreToStackSlot,
652 /// this returns true for any instructions that loads from the
653 /// stack. This is a hint only and may not catch all cases.
654 bool hasStoreToStackSlot(const MachineInstr *MI,
655 const MachineMemOperand *&MMO,
656 int &FrameIndex) const;
658 bool isReallyTriviallyReMaterializable(const MachineInstr *MI,
659 AliasAnalysis *AA) const;
660 void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
661 unsigned DestReg, unsigned SubIdx,
662 const MachineInstr *Orig,
663 const TargetRegisterInfo &TRI) const;
665 /// convertToThreeAddress - This method must be implemented by targets that
666 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
667 /// may be able to convert a two-address instruction into a true
668 /// three-address instruction on demand. This allows the X86 target (for
669 /// example) to convert ADD and SHL instructions into LEA instructions if they
670 /// would require register copies due to two-addressness.
672 /// This method returns a null pointer if the transformation cannot be
673 /// performed, otherwise it returns the new instruction.
675 virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
676 MachineBasicBlock::iterator &MBBI,
677 LiveVariables *LV) const;
679 /// commuteInstruction - We have a few instructions that must be hacked on to
682 virtual MachineInstr *commuteInstruction(MachineInstr *MI, bool NewMI) const;
685 virtual bool isUnpredicatedTerminator(const MachineInstr* MI) const;
686 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
687 MachineBasicBlock *&FBB,
688 SmallVectorImpl<MachineOperand> &Cond,
689 bool AllowModify) const;
690 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const;
691 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
692 MachineBasicBlock *FBB,
693 const SmallVectorImpl<MachineOperand> &Cond,
695 virtual void copyPhysReg(MachineBasicBlock &MBB,
696 MachineBasicBlock::iterator MI, DebugLoc DL,
697 unsigned DestReg, unsigned SrcReg,
699 virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
700 MachineBasicBlock::iterator MI,
701 unsigned SrcReg, bool isKill, int FrameIndex,
702 const TargetRegisterClass *RC,
703 const TargetRegisterInfo *TRI) const;
705 virtual void storeRegToAddr(MachineFunction &MF, unsigned SrcReg, bool isKill,
706 SmallVectorImpl<MachineOperand> &Addr,
707 const TargetRegisterClass *RC,
708 MachineInstr::mmo_iterator MMOBegin,
709 MachineInstr::mmo_iterator MMOEnd,
710 SmallVectorImpl<MachineInstr*> &NewMIs) const;
712 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
713 MachineBasicBlock::iterator MI,
714 unsigned DestReg, int FrameIndex,
715 const TargetRegisterClass *RC,
716 const TargetRegisterInfo *TRI) const;
718 virtual void loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
719 SmallVectorImpl<MachineOperand> &Addr,
720 const TargetRegisterClass *RC,
721 MachineInstr::mmo_iterator MMOBegin,
722 MachineInstr::mmo_iterator MMOEnd,
723 SmallVectorImpl<MachineInstr*> &NewMIs) const;
725 virtual bool spillCalleeSavedRegisters(MachineBasicBlock &MBB,
726 MachineBasicBlock::iterator MI,
727 const std::vector<CalleeSavedInfo> &CSI,
728 const TargetRegisterInfo *TRI) const;
730 virtual bool restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
731 MachineBasicBlock::iterator MI,
732 const std::vector<CalleeSavedInfo> &CSI,
733 const TargetRegisterInfo *TRI) const;
736 MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF,
737 int FrameIx, uint64_t Offset,
741 /// foldMemoryOperand - If this target supports it, fold a load or store of
742 /// the specified stack slot into the specified machine instruction for the
743 /// specified operand(s). If this is possible, the target should perform the
744 /// folding and return true, otherwise it should return false. If it folds
745 /// the instruction, it is likely that the MachineInstruction the iterator
746 /// references has been changed.
747 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
749 const SmallVectorImpl<unsigned> &Ops,
750 int FrameIndex) const;
752 /// foldMemoryOperand - Same as the previous version except it allows folding
753 /// of any load and store from / to any address, not just from a specific
755 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
757 const SmallVectorImpl<unsigned> &Ops,
758 MachineInstr* LoadMI) const;
760 /// canFoldMemoryOperand - Returns true if the specified load / store is
761 /// folding is possible.
762 virtual bool canFoldMemoryOperand(const MachineInstr*,
763 const SmallVectorImpl<unsigned> &) const;
765 /// unfoldMemoryOperand - Separate a single instruction which folded a load or
766 /// a store or a load and a store into two or more instruction. If this is
767 /// possible, returns true as well as the new instructions by reference.
768 virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
769 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
770 SmallVectorImpl<MachineInstr*> &NewMIs) const;
772 virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
773 SmallVectorImpl<SDNode*> &NewNodes) const;
775 /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
776 /// instruction after load / store are unfolded from an instruction of the
777 /// specified opcode. It returns zero if the specified unfolding is not
778 /// possible. If LoadRegIndex is non-null, it is filled in with the operand
779 /// index of the operand which will hold the register holding the loaded
781 virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
782 bool UnfoldLoad, bool UnfoldStore,
783 unsigned *LoadRegIndex = 0) const;
785 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler
786 /// to determine if two loads are loading from the same base address. It
787 /// should only return true if the base pointers are the same and the
788 /// only differences between the two addresses are the offset. It also returns
789 /// the offsets by reference.
790 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
791 int64_t &Offset1, int64_t &Offset2) const;
793 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
794 /// determine (in conjuction with areLoadsFromSameBasePtr) if two loads should
795 /// be scheduled togther. On some targets if two loads are loading from
796 /// addresses in the same cache line, it's better if they are scheduled
797 /// together. This function takes two integers that represent the load offsets
798 /// from the common base address. It returns true if it decides it's desirable
799 /// to schedule the two loads together. "NumLoads" is the number of loads that
800 /// have already been scheduled after Load1.
801 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
802 int64_t Offset1, int64_t Offset2,
803 unsigned NumLoads) const;
805 virtual void getNoopForMachoTarget(MCInst &NopInst) const;
808 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
810 /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine
811 /// instruction that defines the specified register class.
812 bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const;
814 static bool isX86_64NonExtLowByteReg(unsigned reg) {
815 return (reg == X86::SPL || reg == X86::BPL ||
816 reg == X86::SIL || reg == X86::DIL);
819 static bool isX86_64ExtendedReg(const MachineOperand &MO) {
820 if (!MO.isReg()) return false;
821 return isX86_64ExtendedReg(MO.getReg());
823 static unsigned determineREX(const MachineInstr &MI);
825 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended (r8 or
826 /// higher) register? e.g. r8, xmm8, xmm13, etc.
827 static bool isX86_64ExtendedReg(unsigned RegNo);
829 /// GetInstSize - Returns the size of the specified MachineInstr.
831 virtual unsigned GetInstSizeInBytes(const MachineInstr *MI) const;
833 /// getGlobalBaseReg - Return a virtual register initialized with the
834 /// the global base register value. Output instructions required to
835 /// initialize the register in the function entry block, if necessary.
837 unsigned getGlobalBaseReg(MachineFunction *MF) const;
839 /// GetSSEDomain - Return the SSE execution domain of MI as the first element,
840 /// and a bitmask of possible arguments to SetSSEDomain ase the second.
841 std::pair<uint16_t, uint16_t> GetSSEDomain(const MachineInstr *MI) const;
843 /// SetSSEDomain - Set the SSEDomain of MI.
844 void SetSSEDomain(MachineInstr *MI, unsigned Domain) const;
847 MachineInstr * convertToThreeAddressWithLEA(unsigned MIOpc,
848 MachineFunction::iterator &MFI,
849 MachineBasicBlock::iterator &MBBI,
850 LiveVariables *LV) const;
852 MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
855 const SmallVectorImpl<MachineOperand> &MOs,
856 unsigned Size, unsigned Alignment) const;
858 /// isFrameOperand - Return true and the FrameIndex if the specified
859 /// operand and follow operands form a reference to the stack frame.
860 bool isFrameOperand(const MachineInstr *MI, unsigned int Op,
861 int &FrameIndex) const;
864 } // End llvm namespace