class X86TargetMachine;
namespace X86 {
+ // Enums for memory operand decoding. Each memory operand is represented with
+ // a 5 operand sequence in the form:
+ // [BaseReg, ScaleAmt, IndexReg, Disp, Segment]
+ // These enums help decode this.
+ enum {
+ AddrBaseReg = 0,
+ AddrScaleAmt = 1,
+ AddrIndexReg = 2,
+ AddrDisp = 3,
+
+ /// AddrSegmentReg - The operand # of the segment in the memory operand.
+ AddrSegmentReg = 4,
+
+ /// AddrNumOperands - Total number of operands in a memory reference.
+ AddrNumOperands = 5
+ };
+
+
// X86 specific condition code. These correspond to X86_*_COND in
// X86InstrInfo.td. They must be kept in synch.
enum CondCode {
/// MO_DARWIN_STUB - On a symbol operand "FOO", this indicates that the
/// reference is actually to the "FOO$stub" symbol. This is used for calls
- /// and jumps to external functions on Tiger and before.
+ /// and jumps to external functions on Tiger and earlier.
MO_DARWIN_STUB,
/// MO_DARWIN_NONLAZY - On a symbol operand "FOO", this indicates that the
MRM_F8 = 41,
MRM_F9 = 42,
+ /// RawFrmImm8 - This is used for the ENTER instruction, which has two
+ /// immediates, the first of which is a 16-bit immediate (specified by
+ /// the imm encoding) and the second is a 8-bit fixed value.
+ RawFrmImm8 = 43,
+
+ /// RawFrmImm16 - This is used for CALL FAR instructions, which have two
+ /// immediates, the first of which is a 16 or 32-bit immediate (specified by
+ /// the imm encoding) and the second is a 16-bit fixed value. In the AMD
+ /// manual, this operand is described as pntr16:32 and pntr16:16
+ RawFrmImm16 = 44,
+
FormMask = 63,
//===------------------------------------------------------------------===//
Imm8 = 1 << ImmShift,
Imm8PCRel = 2 << ImmShift,
Imm16 = 3 << ImmShift,
- Imm32 = 4 << ImmShift,
- Imm32PCRel = 5 << ImmShift,
- Imm64 = 6 << ImmShift,
+ Imm16PCRel = 4 << ImmShift,
+ Imm32 = 5 << ImmShift,
+ Imm32PCRel = 6 << ImmShift,
+ Imm64 = 7 << ImmShift,
//===------------------------------------------------------------------===//
// FP Instruction Classification... Zero is non-fp instruction.
SSEDomainShift = 22,
OpcodeShift = 24,
- OpcodeMask = 0xFF << OpcodeShift
+ OpcodeMask = 0xFF << OpcodeShift,
- };
-
- // FIXME: The enum opcode space is over and more bits are needed. Anywhere
- // those enums below are used, TSFlags must be shifted right by 32 first.
- enum {
//===------------------------------------------------------------------===//
- // VEX_4V - VEX prefixes are instruction prefixes used in AVX.
- // VEX_4V is used to specify an additional AVX/SSE register. Several 2
- // address instructions in SSE are represented as 3 address ones in AVX
- // and the additional register is encoded in VEX_VVVV prefix.
- //
- VEXShift = 0,
- VEX_4V = 1 << VEXShift
+ /// VEX - The opcode prefix used by AVX instructions
+ VEX = 1U << 0,
+
+ /// VEX_W - Has a opcode specific functionality, but is used in the same
+ /// way as REX_W is for regular SSE instructions.
+ VEX_W = 1U << 1,
+
+ /// VEX_4V - Used to specify an additional AVX/SSE register. Several 2
+ /// address instructions in SSE are represented as 3 address ones in AVX
+ /// and the additional register is encoded in VEX_VVVV prefix.
+ VEX_4V = 1U << 2,
+
+ /// VEX_I8IMM - Specifies that the last register used in a AVX instruction,
+ /// must be encoded in the i8 immediate field. This usually happens in
+ /// instructions with 4 operands.
+ VEX_I8IMM = 1U << 3,
+
+ /// VEX_L - Stands for a bit in the VEX opcode prefix meaning the current
+ /// instruction uses 256-bit wide registers. This is usually auto detected
+ /// if a VR256 register is used, but some AVX instructions also have this
+ /// field marked when using a f256 memory references.
+ VEX_L = 1U << 4,
+
+ /// Has3DNow0F0FOpcode - This flag indicates that the instruction uses the
+ /// wacky 0x0F 0x0F prefix for 3DNow! instructions. The manual documents
+ /// this as having a 0x0F prefix with a 0x0F opcode, and each instruction
+ /// storing a classifier in the imm8 field. To simplify our implementation,
+ /// we handle this by storeing the classifier in the opcode field and using
+ /// this flag to indicate that the encoder should do the wacky 3DNow! thing.
+ Has3DNow0F0FOpcode = 1U << 5
};
-
+
// getBaseOpcodeFor - This function returns the "base" X86 opcode for the
// specified machine instruction.
//
default: assert(0 && "Unknown immediate size");
case X86II::Imm8:
case X86II::Imm8PCRel: return 1;
- case X86II::Imm16: return 2;
+ case X86II::Imm16:
+ case X86II::Imm16PCRel: return 2;
case X86II::Imm32:
case X86II::Imm32PCRel: return 4;
case X86II::Imm64: return 8;
/// TSFlags indicates that it is pc relative.
static inline unsigned isImmPCRel(uint64_t TSFlags) {
switch (TSFlags & X86II::ImmMask) {
- default: assert(0 && "Unknown immediate size");
- case X86II::Imm8PCRel:
- case X86II::Imm32PCRel:
- return true;
- case X86II::Imm8:
- case X86II::Imm16:
- case X86II::Imm32:
- case X86II::Imm64:
- return false;
+ default: assert(0 && "Unknown immediate size");
+ case X86II::Imm8PCRel:
+ case X86II::Imm16PCRel:
+ case X86II::Imm32PCRel:
+ return true;
+ case X86II::Imm8:
+ case X86II::Imm16:
+ case X86II::Imm32:
+ case X86II::Imm64:
+ return false;
}
- }
+ }
+
+ /// getMemoryOperandNo - The function returns the MCInst operand # for the
+ /// first field of the memory operand. If the instruction doesn't have a
+ /// memory operand, this returns -1.
+ ///
+ /// Note that this ignores tied operands. If there is a tied register which
+ /// is duplicated in the MCInst (e.g. "EAX = addl EAX, [mem]") it is only
+ /// counted as one operand.
+ ///
+ static inline int getMemoryOperandNo(uint64_t TSFlags) {
+ switch (TSFlags & X86II::FormMask) {
+ case X86II::MRMInitReg: assert(0 && "FIXME: Remove this form");
+ default: assert(0 && "Unknown FormMask value in getMemoryOperandNo!");
+ case X86II::Pseudo:
+ case X86II::RawFrm:
+ case X86II::AddRegFrm:
+ case X86II::MRMDestReg:
+ case X86II::MRMSrcReg:
+ case X86II::RawFrmImm8:
+ case X86II::RawFrmImm16:
+ return -1;
+ case X86II::MRMDestMem:
+ return 0;
+ case X86II::MRMSrcMem: {
+ bool HasVEX_4V = (TSFlags >> 32) & X86II::VEX_4V;
+ unsigned FirstMemOp = 1;
+ if (HasVEX_4V)
+ ++FirstMemOp;// Skip the register source (which is encoded in VEX_VVVV).
+
+ // FIXME: Maybe lea should have its own form? This is a horrible hack.
+ //if (Opcode == X86::LEA64r || Opcode == X86::LEA64_32r ||
+ // Opcode == X86::LEA16r || Opcode == X86::LEA32r)
+ return FirstMemOp;
+ }
+ case X86II::MRM0r: case X86II::MRM1r:
+ case X86II::MRM2r: case X86II::MRM3r:
+ case X86II::MRM4r: case X86II::MRM5r:
+ case X86II::MRM6r: case X86II::MRM7r:
+ return -1;
+ case X86II::MRM0m: case X86II::MRM1m:
+ case X86II::MRM2m: case X86II::MRM3m:
+ case X86II::MRM4m: case X86II::MRM5m:
+ case X86II::MRM6m: case X86II::MRM7m:
+ return 0;
+ case X86II::MRM_C1:
+ case X86II::MRM_C2:
+ case X86II::MRM_C3:
+ case X86II::MRM_C4:
+ case X86II::MRM_C8:
+ case X86II::MRM_C9:
+ case X86II::MRM_E8:
+ case X86II::MRM_F0:
+ case X86II::MRM_F8:
+ case X86II::MRM_F9:
+ return -1;
+ }
+ }
}
-const int X86AddrNumOperands = 5;
-
inline static bool isScale(const MachineOperand &MO) {
return MO.isImm() &&
(MO.getImm() == 1 || MO.getImm() == 2 ||
/// RegOp2MemOpTable2Addr, RegOp2MemOpTable0, RegOp2MemOpTable1,
/// RegOp2MemOpTable2 - Load / store folding opcode maps.
///
- DenseMap<unsigned*, std::pair<unsigned,unsigned> > RegOp2MemOpTable2Addr;
- DenseMap<unsigned*, std::pair<unsigned,unsigned> > RegOp2MemOpTable0;
- DenseMap<unsigned*, std::pair<unsigned,unsigned> > RegOp2MemOpTable1;
- DenseMap<unsigned*, std::pair<unsigned,unsigned> > RegOp2MemOpTable2;
+ DenseMap<unsigned, std::pair<unsigned,unsigned> > RegOp2MemOpTable2Addr;
+ DenseMap<unsigned, std::pair<unsigned,unsigned> > RegOp2MemOpTable0;
+ DenseMap<unsigned, std::pair<unsigned,unsigned> > RegOp2MemOpTable1;
+ DenseMap<unsigned, std::pair<unsigned,unsigned> > RegOp2MemOpTable2;
/// MemOp2RegOpTable - Load / store unfolding opcode map.
///
- DenseMap<unsigned*, std::pair<unsigned, unsigned> > MemOp2RegOpTable;
+ DenseMap<unsigned, std::pair<unsigned, unsigned> > MemOp2RegOpTable;
public:
explicit X86InstrInfo(X86TargetMachine &tm);
///
virtual const X86RegisterInfo &getRegisterInfo() const { return RI; }
- /// Return true if the instruction is a register to register move and return
- /// the source and dest operands and their sub-register indices by reference.
- virtual bool isMoveInstr(const MachineInstr &MI,
- unsigned &SrcReg, unsigned &DstReg,
- unsigned &SrcSubIdx, unsigned &DstSubIdx) const;
-
/// isCoalescableExtInstr - Return true if the instruction is a "coalescable"
/// extension instruction. That is, it's like a copy where it's legal for the
/// source to overlap the destination. e.g. X86::MOVSX64rr32. If this returns
MachineBasicBlock *FBB,
const SmallVectorImpl<MachineOperand> &Cond,
DebugLoc DL) const;
- virtual bool copyRegToReg(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MI,
- unsigned DestReg, unsigned SrcReg,
- const TargetRegisterClass *DestRC,
- const TargetRegisterClass *SrcRC,
- DebugLoc DL) const;
+ virtual void copyPhysReg(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator MI, DebugLoc DL,
+ unsigned DestReg, unsigned SrcReg,
+ bool KillSrc) const;
virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI,
unsigned SrcReg, bool isKill, int FrameIndex,
MachineInstr::mmo_iterator MMOBegin,
MachineInstr::mmo_iterator MMOEnd,
SmallVectorImpl<MachineInstr*> &NewMIs) const;
-
- virtual bool spillCalleeSavedRegisters(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MI,
- const std::vector<CalleeSavedInfo> &CSI,
- const TargetRegisterInfo *TRI) const;
-
- virtual bool restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MI,
- const std::vector<CalleeSavedInfo> &CSI,
- const TargetRegisterInfo *TRI) const;
-
virtual
MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF,
int FrameIx, uint64_t Offset,
if (!MO.isReg()) return false;
return isX86_64ExtendedReg(MO.getReg());
}
- static unsigned determineREX(const MachineInstr &MI);
/// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended (r8 or
/// higher) register? e.g. r8, xmm8, xmm13, etc.
static bool isX86_64ExtendedReg(unsigned RegNo);
- /// GetInstSize - Returns the size of the specified MachineInstr.
- ///
- virtual unsigned GetInstSizeInBytes(const MachineInstr *MI) const;
-
/// getGlobalBaseReg - Return a virtual register initialized with the
/// the global base register value. Output instructions required to
/// initialize the register in the function entry block, if necessary.
/// SetSSEDomain - Set the SSEDomain of MI.
void SetSSEDomain(MachineInstr *MI, unsigned Domain) const;
+ MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
+ MachineInstr* MI,
+ unsigned OpNum,
+ const SmallVectorImpl<MachineOperand> &MOs,
+ unsigned Size, unsigned Alignment) const;
+
+ bool hasHighOperandLatency(const InstrItineraryData *ItinData,
+ const MachineRegisterInfo *MRI,
+ const MachineInstr *DefMI, unsigned DefIdx,
+ const MachineInstr *UseMI, unsigned UseIdx) const;
+
private:
MachineInstr * convertToThreeAddressWithLEA(unsigned MIOpc,
MachineFunction::iterator &MFI,
MachineBasicBlock::iterator &MBBI,
LiveVariables *LV) const;
- MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
- MachineInstr* MI,
- unsigned OpNum,
- const SmallVectorImpl<MachineOperand> &MOs,
- unsigned Size, unsigned Alignment) const;
-
/// isFrameOperand - Return true and the FrameIndex if the specified
/// operand and follow operands form a reference to the stack frame.
bool isFrameOperand(const MachineInstr *MI, unsigned int Op,