1 //===-- X86/X86MCCodeEmitter.cpp - Convert X86 code to machine code -------===//
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 implements the X86MCCodeEmitter class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "x86-emitter"
16 #include "X86InstrInfo.h"
17 #include "X86FixupKinds.h"
18 #include "llvm/MC/MCCodeEmitter.h"
19 #include "llvm/MC/MCExpr.h"
20 #include "llvm/MC/MCInst.h"
21 #include "llvm/Support/raw_ostream.h"
25 class X86MCCodeEmitter : public MCCodeEmitter {
26 X86MCCodeEmitter(const X86MCCodeEmitter &); // DO NOT IMPLEMENT
27 void operator=(const X86MCCodeEmitter &); // DO NOT IMPLEMENT
28 const TargetMachine &TM;
29 const TargetInstrInfo &TII;
33 X86MCCodeEmitter(TargetMachine &tm, MCContext &ctx, bool is64Bit)
34 : TM(tm), TII(*TM.getInstrInfo()), Ctx(ctx) {
35 Is64BitMode = is64Bit;
38 ~X86MCCodeEmitter() {}
40 unsigned getNumFixupKinds() const {
44 const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const {
45 const static MCFixupKindInfo Infos[] = {
46 { "reloc_pcrel_4byte", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel },
47 { "reloc_pcrel_1byte", 0, 1 * 8, MCFixupKindInfo::FKF_IsPCRel },
48 { "reloc_pcrel_2byte", 0, 2 * 8, MCFixupKindInfo::FKF_IsPCRel },
49 { "reloc_riprel_4byte", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel },
50 { "reloc_riprel_4byte_movq_load", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel },
51 { "reloc_signed_4byte", 0, 4 * 8, 0}
54 if (Kind < FirstTargetFixupKind)
55 return MCCodeEmitter::getFixupKindInfo(Kind);
57 assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
59 return Infos[Kind - FirstTargetFixupKind];
62 static unsigned GetX86RegNum(const MCOperand &MO) {
63 return X86RegisterInfo::getX86RegNum(MO.getReg());
66 // On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the range
67 // 0-7 and the difference between the 2 groups is given by the REX prefix.
68 // In the VEX prefix, registers are seen sequencially from 0-15 and encoded
69 // in 1's complement form, example:
71 // ModRM field => XMM9 => 1
72 // VEX.VVVV => XMM9 => ~9
74 // See table 4-35 of Intel AVX Programming Reference for details.
75 static unsigned char getVEXRegisterEncoding(const MCInst &MI,
77 unsigned SrcReg = MI.getOperand(OpNum).getReg();
78 unsigned SrcRegNum = GetX86RegNum(MI.getOperand(OpNum));
79 if ((SrcReg >= X86::XMM8 && SrcReg <= X86::XMM15) ||
80 (SrcReg >= X86::YMM8 && SrcReg <= X86::YMM15))
83 // The registers represented through VEX_VVVV should
84 // be encoded in 1's complement form.
85 return (~SrcRegNum) & 0xf;
88 void EmitByte(unsigned char C, unsigned &CurByte, raw_ostream &OS) const {
93 void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
94 raw_ostream &OS) const {
95 // Output the constant in little endian byte order.
96 for (unsigned i = 0; i != Size; ++i) {
97 EmitByte(Val & 255, CurByte, OS);
102 void EmitImmediate(const MCOperand &Disp,
103 unsigned ImmSize, MCFixupKind FixupKind,
104 unsigned &CurByte, raw_ostream &OS,
105 SmallVectorImpl<MCFixup> &Fixups,
106 int ImmOffset = 0) const;
108 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
110 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
111 return RM | (RegOpcode << 3) | (Mod << 6);
114 void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
115 unsigned &CurByte, raw_ostream &OS) const {
116 EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS);
119 void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base,
120 unsigned &CurByte, raw_ostream &OS) const {
121 // SIB byte is in the same format as the ModRMByte.
122 EmitByte(ModRMByte(SS, Index, Base), CurByte, OS);
126 void EmitMemModRMByte(const MCInst &MI, unsigned Op,
127 unsigned RegOpcodeField,
128 uint64_t TSFlags, unsigned &CurByte, raw_ostream &OS,
129 SmallVectorImpl<MCFixup> &Fixups) const;
131 void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
132 SmallVectorImpl<MCFixup> &Fixups) const;
134 void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
135 const MCInst &MI, const TargetInstrDesc &Desc,
136 raw_ostream &OS) const;
138 void EmitSegmentOverridePrefix(uint64_t TSFlags, unsigned &CurByte,
139 int MemOperand, const MCInst &MI,
140 raw_ostream &OS) const;
142 void EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
143 const MCInst &MI, const TargetInstrDesc &Desc,
144 raw_ostream &OS) const;
147 } // end anonymous namespace
150 MCCodeEmitter *llvm::createX86_32MCCodeEmitter(const Target &,
153 return new X86MCCodeEmitter(TM, Ctx, false);
156 MCCodeEmitter *llvm::createX86_64MCCodeEmitter(const Target &,
159 return new X86MCCodeEmitter(TM, Ctx, true);
162 /// isDisp8 - Return true if this signed displacement fits in a 8-bit
163 /// sign-extended field.
164 static bool isDisp8(int Value) {
165 return Value == (signed char)Value;
168 /// getImmFixupKind - Return the appropriate fixup kind to use for an immediate
169 /// in an instruction with the specified TSFlags.
170 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
171 unsigned Size = X86II::getSizeOfImm(TSFlags);
172 bool isPCRel = X86II::isImmPCRel(TSFlags);
175 default: assert(0 && "Unknown immediate size");
176 case 1: return isPCRel ? MCFixupKind(X86::reloc_pcrel_1byte) : FK_Data_1;
177 case 2: return isPCRel ? MCFixupKind(X86::reloc_pcrel_2byte) : FK_Data_2;
178 case 4: return isPCRel ? MCFixupKind(X86::reloc_pcrel_4byte) : FK_Data_4;
179 case 8: assert(!isPCRel); return FK_Data_8;
183 /// Is32BitMemOperand - Return true if the specified instruction with a memory
184 /// operand should emit the 0x67 prefix byte in 64-bit mode due to a 32-bit
185 /// memory operand. Op specifies the operand # of the memoperand.
186 static bool Is32BitMemOperand(const MCInst &MI, unsigned Op) {
187 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
188 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
190 if ((BaseReg.getReg() != 0 && X86::GR32RegClass.contains(BaseReg.getReg())) ||
191 (IndexReg.getReg() != 0 && X86::GR32RegClass.contains(IndexReg.getReg())))
196 void X86MCCodeEmitter::
197 EmitImmediate(const MCOperand &DispOp, unsigned Size, MCFixupKind FixupKind,
198 unsigned &CurByte, raw_ostream &OS,
199 SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
200 // If this is a simple integer displacement that doesn't require a relocation,
202 if (DispOp.isImm()) {
203 // FIXME: is this right for pc-rel encoding?? Probably need to emit this as
205 EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
209 // If we have an immoffset, add it to the expression.
210 const MCExpr *Expr = DispOp.getExpr();
212 // If the fixup is pc-relative, we need to bias the value to be relative to
213 // the start of the field, not the end of the field.
214 if (FixupKind == MCFixupKind(X86::reloc_pcrel_4byte) ||
215 FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
216 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load))
218 if (FixupKind == MCFixupKind(X86::reloc_pcrel_2byte))
220 if (FixupKind == MCFixupKind(X86::reloc_pcrel_1byte))
224 Expr = MCBinaryExpr::CreateAdd(Expr, MCConstantExpr::Create(ImmOffset, Ctx),
227 // Emit a symbolic constant as a fixup and 4 zeros.
228 Fixups.push_back(MCFixup::Create(CurByte, Expr, FixupKind));
229 EmitConstant(0, Size, CurByte, OS);
232 void X86MCCodeEmitter::EmitMemModRMByte(const MCInst &MI, unsigned Op,
233 unsigned RegOpcodeField,
234 uint64_t TSFlags, unsigned &CurByte,
236 SmallVectorImpl<MCFixup> &Fixups) const{
237 const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
238 const MCOperand &Base = MI.getOperand(Op+X86::AddrBaseReg);
239 const MCOperand &Scale = MI.getOperand(Op+X86::AddrScaleAmt);
240 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
241 unsigned BaseReg = Base.getReg();
243 // Handle %rip relative addressing.
244 if (BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
245 assert(Is64BitMode && "Rip-relative addressing requires 64-bit mode");
246 assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
247 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
249 unsigned FixupKind = X86::reloc_riprel_4byte;
251 // movq loads are handled with a special relocation form which allows the
252 // linker to eliminate some loads for GOT references which end up in the
253 // same linkage unit.
254 if (MI.getOpcode() == X86::MOV64rm)
255 FixupKind = X86::reloc_riprel_4byte_movq_load;
257 // rip-relative addressing is actually relative to the *next* instruction.
258 // Since an immediate can follow the mod/rm byte for an instruction, this
259 // means that we need to bias the immediate field of the instruction with
260 // the size of the immediate field. If we have this case, add it into the
261 // expression to emit.
262 int ImmSize = X86II::hasImm(TSFlags) ? X86II::getSizeOfImm(TSFlags) : 0;
264 EmitImmediate(Disp, 4, MCFixupKind(FixupKind),
265 CurByte, OS, Fixups, -ImmSize);
269 unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
271 // Determine whether a SIB byte is needed.
272 // If no BaseReg, issue a RIP relative instruction only if the MCE can
273 // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
274 // 2-7) and absolute references.
276 if (// The SIB byte must be used if there is an index register.
277 IndexReg.getReg() == 0 &&
278 // The SIB byte must be used if the base is ESP/RSP/R12, all of which
279 // encode to an R/M value of 4, which indicates that a SIB byte is
281 BaseRegNo != N86::ESP &&
282 // If there is no base register and we're in 64-bit mode, we need a SIB
283 // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
284 (!Is64BitMode || BaseReg != 0)) {
286 if (BaseReg == 0) { // [disp32] in X86-32 mode
287 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
288 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
292 // If the base is not EBP/ESP and there is no displacement, use simple
293 // indirect register encoding, this handles addresses like [EAX]. The
294 // encoding for [EBP] with no displacement means [disp32] so we handle it
295 // by emitting a displacement of 0 below.
296 if (Disp.isImm() && Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
297 EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
301 // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
302 if (Disp.isImm() && isDisp8(Disp.getImm())) {
303 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
304 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
308 // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
309 EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
310 EmitImmediate(Disp, 4, MCFixupKind(X86::reloc_signed_4byte), CurByte, OS,
315 // We need a SIB byte, so start by outputting the ModR/M byte first
316 assert(IndexReg.getReg() != X86::ESP &&
317 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
319 bool ForceDisp32 = false;
320 bool ForceDisp8 = false;
322 // If there is no base register, we emit the special case SIB byte with
323 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
324 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
326 } else if (!Disp.isImm()) {
327 // Emit the normal disp32 encoding.
328 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
330 } else if (Disp.getImm() == 0 &&
331 // Base reg can't be anything that ends up with '5' as the base
332 // reg, it is the magic [*] nomenclature that indicates no base.
333 BaseRegNo != N86::EBP) {
334 // Emit no displacement ModR/M byte
335 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
336 } else if (isDisp8(Disp.getImm())) {
337 // Emit the disp8 encoding.
338 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
339 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
341 // Emit the normal disp32 encoding.
342 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
345 // Calculate what the SS field value should be...
346 static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 };
347 unsigned SS = SSTable[Scale.getImm()];
350 // Handle the SIB byte for the case where there is no base, see Intel
351 // Manual 2A, table 2-7. The displacement has already been output.
353 if (IndexReg.getReg())
354 IndexRegNo = GetX86RegNum(IndexReg);
355 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
357 EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
360 if (IndexReg.getReg())
361 IndexRegNo = GetX86RegNum(IndexReg);
363 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
364 EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
367 // Do we need to output a displacement?
369 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
370 else if (ForceDisp32 || Disp.getImm() != 0)
371 EmitImmediate(Disp, 4, MCFixupKind(X86::reloc_signed_4byte), CurByte, OS,
375 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
377 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
378 int MemOperand, const MCInst &MI,
379 const TargetInstrDesc &Desc,
380 raw_ostream &OS) const {
381 bool HasVEX_4V = false;
382 if ((TSFlags >> 32) & X86II::VEX_4V)
385 // VEX_R: opcode externsion equivalent to REX.R in
386 // 1's complement (inverted) form
388 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
389 // 0: Same as REX_R=1 (64 bit mode only)
391 unsigned char VEX_R = 0x1;
393 // VEX_X: equivalent to REX.X, only used when a
394 // register is used for index in SIB Byte.
396 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
397 // 0: Same as REX.X=1 (64-bit mode only)
398 unsigned char VEX_X = 0x1;
402 // 1: Same as REX_B=0 (ignored in 32-bit mode)
403 // 0: Same as REX_B=1 (64 bit mode only)
405 unsigned char VEX_B = 0x1;
407 // VEX_W: opcode specific (use like REX.W, or used for
408 // opcode extension, or ignored, depending on the opcode byte)
409 unsigned char VEX_W = 0;
411 // VEX_5M (VEX m-mmmmm field):
413 // 0b00000: Reserved for future use
414 // 0b00001: implied 0F leading opcode
415 // 0b00010: implied 0F 38 leading opcode bytes
416 // 0b00011: implied 0F 3A leading opcode bytes
417 // 0b00100-0b11111: Reserved for future use
419 unsigned char VEX_5M = 0x1;
421 // VEX_4V (VEX vvvv field): a register specifier
422 // (in 1's complement form) or 1111 if unused.
423 unsigned char VEX_4V = 0xf;
425 // VEX_L (Vector Length):
427 // 0: scalar or 128-bit vector
430 unsigned char VEX_L = 0;
432 // VEX_PP: opcode extension providing equivalent
433 // functionality of a SIMD prefix
440 unsigned char VEX_PP = 0;
442 // Encode the operand size opcode prefix as needed.
443 if (TSFlags & X86II::OpSize)
446 if ((TSFlags >> 32) & X86II::VEX_W)
449 if ((TSFlags >> 32) & X86II::VEX_L)
452 switch (TSFlags & X86II::Op0Mask) {
453 default: assert(0 && "Invalid prefix!");
454 case X86II::T8: // 0F 38
457 case X86II::TA: // 0F 3A
460 case X86II::TF: // F2 0F 38
464 case X86II::XS: // F3 0F
467 case X86II::XD: // F2 0F
470 case X86II::TB: // Bypass: Not used by VEX
475 // Set the vector length to 256-bit if YMM0-YMM15 is used
476 for (unsigned i = 0; i != MI.getNumOperands(); ++i) {
477 if (!MI.getOperand(i).isReg())
479 unsigned SrcReg = MI.getOperand(i).getReg();
480 if (SrcReg >= X86::YMM0 && SrcReg <= X86::YMM15)
484 unsigned NumOps = MI.getNumOperands();
486 bool IsDestMem = false;
488 switch (TSFlags & X86II::FormMask) {
489 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
490 case X86II::MRMDestMem:
492 // The important info for the VEX prefix is never beyond the address
493 // registers. Don't check beyond that.
494 NumOps = CurOp = X86::AddrNumOperands;
495 case X86II::MRM0m: case X86II::MRM1m:
496 case X86II::MRM2m: case X86II::MRM3m:
497 case X86II::MRM4m: case X86II::MRM5m:
498 case X86II::MRM6m: case X86II::MRM7m:
499 case X86II::MRMSrcMem:
500 case X86II::MRMSrcReg:
501 if (MI.getNumOperands() > CurOp && MI.getOperand(CurOp).isReg() &&
502 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
507 VEX_4V = getVEXRegisterEncoding(MI, IsDestMem ? CurOp-1 : CurOp);
511 // To only check operands before the memory address ones, start
512 // the search from the begining
516 // If the last register should be encoded in the immediate field
517 // do not use any bit from VEX prefix to this register, ignore it
518 if ((TSFlags >> 32) & X86II::VEX_I8IMM)
521 for (; CurOp != NumOps; ++CurOp) {
522 const MCOperand &MO = MI.getOperand(CurOp);
523 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
525 if (!VEX_B && MO.isReg() &&
526 ((TSFlags & X86II::FormMask) == X86II::MRMSrcMem) &&
527 X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
531 default: // MRMDestReg, MRM0r-MRM7r, RawFrm
532 if (!MI.getNumOperands())
535 if (MI.getOperand(CurOp).isReg() &&
536 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
540 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
543 for (; CurOp != NumOps; ++CurOp) {
544 const MCOperand &MO = MI.getOperand(CurOp);
545 if (MO.isReg() && !HasVEX_4V &&
546 X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
552 // Emit segment override opcode prefix as needed.
553 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
555 // VEX opcode prefix can have 2 or 3 bytes
558 // +-----+ +--------------+ +-------------------+
559 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
560 // +-----+ +--------------+ +-------------------+
562 // +-----+ +-------------------+
563 // | C5h | | R | vvvv | L | pp |
564 // +-----+ +-------------------+
566 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
568 if (VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) { // 2 byte VEX prefix
569 EmitByte(0xC5, CurByte, OS);
570 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
575 EmitByte(0xC4, CurByte, OS);
576 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
577 EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
580 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
581 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
582 /// size, and 3) use of X86-64 extended registers.
583 static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
584 const TargetInstrDesc &Desc) {
586 if (TSFlags & X86II::REX_W)
587 REX |= 1 << 3; // set REX.W
589 if (MI.getNumOperands() == 0) return REX;
591 unsigned NumOps = MI.getNumOperands();
592 // FIXME: MCInst should explicitize the two-addrness.
593 bool isTwoAddr = NumOps > 1 &&
594 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
596 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
597 unsigned i = isTwoAddr ? 1 : 0;
598 for (; i != NumOps; ++i) {
599 const MCOperand &MO = MI.getOperand(i);
600 if (!MO.isReg()) continue;
601 unsigned Reg = MO.getReg();
602 if (!X86InstrInfo::isX86_64NonExtLowByteReg(Reg)) continue;
603 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
604 // that returns non-zero.
605 REX |= 0x40; // REX fixed encoding prefix
609 switch (TSFlags & X86II::FormMask) {
610 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
611 case X86II::MRMSrcReg:
612 if (MI.getOperand(0).isReg() &&
613 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
614 REX |= 1 << 2; // set REX.R
615 i = isTwoAddr ? 2 : 1;
616 for (; i != NumOps; ++i) {
617 const MCOperand &MO = MI.getOperand(i);
618 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
619 REX |= 1 << 0; // set REX.B
622 case X86II::MRMSrcMem: {
623 if (MI.getOperand(0).isReg() &&
624 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
625 REX |= 1 << 2; // set REX.R
627 i = isTwoAddr ? 2 : 1;
628 for (; i != NumOps; ++i) {
629 const MCOperand &MO = MI.getOperand(i);
631 if (X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
632 REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1)
638 case X86II::MRM0m: case X86II::MRM1m:
639 case X86II::MRM2m: case X86II::MRM3m:
640 case X86II::MRM4m: case X86II::MRM5m:
641 case X86II::MRM6m: case X86II::MRM7m:
642 case X86II::MRMDestMem: {
643 unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands);
644 i = isTwoAddr ? 1 : 0;
645 if (NumOps > e && MI.getOperand(e).isReg() &&
646 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(e).getReg()))
647 REX |= 1 << 2; // set REX.R
649 for (; i != e; ++i) {
650 const MCOperand &MO = MI.getOperand(i);
652 if (X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
653 REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1)
660 if (MI.getOperand(0).isReg() &&
661 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
662 REX |= 1 << 0; // set REX.B
663 i = isTwoAddr ? 2 : 1;
664 for (unsigned e = NumOps; i != e; ++i) {
665 const MCOperand &MO = MI.getOperand(i);
666 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
667 REX |= 1 << 2; // set REX.R
674 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
675 void X86MCCodeEmitter::EmitSegmentOverridePrefix(uint64_t TSFlags,
676 unsigned &CurByte, int MemOperand,
678 raw_ostream &OS) const {
679 switch (TSFlags & X86II::SegOvrMask) {
680 default: assert(0 && "Invalid segment!");
682 // No segment override, check for explicit one on memory operand.
683 if (MemOperand != -1) { // If the instruction has a memory operand.
684 switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
685 default: assert(0 && "Unknown segment register!");
687 case X86::CS: EmitByte(0x2E, CurByte, OS); break;
688 case X86::SS: EmitByte(0x36, CurByte, OS); break;
689 case X86::DS: EmitByte(0x3E, CurByte, OS); break;
690 case X86::ES: EmitByte(0x26, CurByte, OS); break;
691 case X86::FS: EmitByte(0x64, CurByte, OS); break;
692 case X86::GS: EmitByte(0x65, CurByte, OS); break;
697 EmitByte(0x64, CurByte, OS);
700 EmitByte(0x65, CurByte, OS);
705 /// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode.
707 /// MemOperand is the operand # of the start of a memory operand if present. If
708 /// Not present, it is -1.
709 void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
710 int MemOperand, const MCInst &MI,
711 const TargetInstrDesc &Desc,
712 raw_ostream &OS) const {
714 // Emit the lock opcode prefix as needed.
715 if (TSFlags & X86II::LOCK)
716 EmitByte(0xF0, CurByte, OS);
718 // Emit segment override opcode prefix as needed.
719 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
721 // Emit the repeat opcode prefix as needed.
722 if ((TSFlags & X86II::Op0Mask) == X86II::REP)
723 EmitByte(0xF3, CurByte, OS);
725 // Emit the address size opcode prefix as needed.
726 if ((TSFlags & X86II::AdSize) ||
727 (MemOperand != -1 && Is64BitMode && Is32BitMemOperand(MI, MemOperand)))
728 EmitByte(0x67, CurByte, OS);
730 // Emit the operand size opcode prefix as needed.
731 if (TSFlags & X86II::OpSize)
732 EmitByte(0x66, CurByte, OS);
734 bool Need0FPrefix = false;
735 switch (TSFlags & X86II::Op0Mask) {
736 default: assert(0 && "Invalid prefix!");
737 case 0: break; // No prefix!
738 case X86II::REP: break; // already handled.
739 case X86II::TB: // Two-byte opcode prefix
740 case X86II::T8: // 0F 38
741 case X86II::TA: // 0F 3A
744 case X86II::TF: // F2 0F 38
745 EmitByte(0xF2, CurByte, OS);
748 case X86II::XS: // F3 0F
749 EmitByte(0xF3, CurByte, OS);
752 case X86II::XD: // F2 0F
753 EmitByte(0xF2, CurByte, OS);
756 case X86II::D8: EmitByte(0xD8, CurByte, OS); break;
757 case X86II::D9: EmitByte(0xD9, CurByte, OS); break;
758 case X86II::DA: EmitByte(0xDA, CurByte, OS); break;
759 case X86II::DB: EmitByte(0xDB, CurByte, OS); break;
760 case X86II::DC: EmitByte(0xDC, CurByte, OS); break;
761 case X86II::DD: EmitByte(0xDD, CurByte, OS); break;
762 case X86II::DE: EmitByte(0xDE, CurByte, OS); break;
763 case X86II::DF: EmitByte(0xDF, CurByte, OS); break;
766 // Handle REX prefix.
767 // FIXME: Can this come before F2 etc to simplify emission?
769 if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc))
770 EmitByte(0x40 | REX, CurByte, OS);
773 // 0x0F escape code must be emitted just before the opcode.
775 EmitByte(0x0F, CurByte, OS);
777 // FIXME: Pull this up into previous switch if REX can be moved earlier.
778 switch (TSFlags & X86II::Op0Mask) {
779 case X86II::TF: // F2 0F 38
780 case X86II::T8: // 0F 38
781 EmitByte(0x38, CurByte, OS);
783 case X86II::TA: // 0F 3A
784 EmitByte(0x3A, CurByte, OS);
789 void X86MCCodeEmitter::
790 EncodeInstruction(const MCInst &MI, raw_ostream &OS,
791 SmallVectorImpl<MCFixup> &Fixups) const {
792 unsigned Opcode = MI.getOpcode();
793 const TargetInstrDesc &Desc = TII.get(Opcode);
794 uint64_t TSFlags = Desc.TSFlags;
796 // Pseudo instructions don't get encoded.
797 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
800 // If this is a two-address instruction, skip one of the register operands.
801 // FIXME: This should be handled during MCInst lowering.
802 unsigned NumOps = Desc.getNumOperands();
804 if (NumOps > 1 && Desc.getOperandConstraint(1, TOI::TIED_TO) != -1)
806 else if (NumOps > 2 && Desc.getOperandConstraint(NumOps-1, TOI::TIED_TO)== 0)
807 // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32
810 // Keep track of the current byte being emitted.
811 unsigned CurByte = 0;
813 // Is this instruction encoded using the AVX VEX prefix?
814 bool HasVEXPrefix = false;
816 // It uses the VEX.VVVV field?
817 bool HasVEX_4V = false;
819 if ((TSFlags >> 32) & X86II::VEX)
821 if ((TSFlags >> 32) & X86II::VEX_4V)
825 // Determine where the memory operand starts, if present.
826 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags);
827 if (MemoryOperand != -1) MemoryOperand += CurOp;
830 EmitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
832 EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
835 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
837 if ((TSFlags >> 32) & X86II::Has3DNow0F0FOpcode)
838 BaseOpcode = 0x0F; // Weird 3DNow! encoding.
840 unsigned SrcRegNum = 0;
841 switch (TSFlags & X86II::FormMask) {
842 case X86II::MRMInitReg:
843 assert(0 && "FIXME: Remove this form when the JIT moves to MCCodeEmitter!");
844 default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
845 assert(0 && "Unknown FormMask value in X86MCCodeEmitter!");
847 assert(0 && "Pseudo instruction shouldn't be emitted");
849 EmitByte(BaseOpcode, CurByte, OS);
852 case X86II::RawFrmImm8:
853 EmitByte(BaseOpcode, CurByte, OS);
854 EmitImmediate(MI.getOperand(CurOp++),
855 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
856 CurByte, OS, Fixups);
857 EmitImmediate(MI.getOperand(CurOp++), 1, FK_Data_1, CurByte, OS, Fixups);
859 case X86II::RawFrmImm16:
860 EmitByte(BaseOpcode, CurByte, OS);
861 EmitImmediate(MI.getOperand(CurOp++),
862 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
863 CurByte, OS, Fixups);
864 EmitImmediate(MI.getOperand(CurOp++), 2, FK_Data_2, CurByte, OS, Fixups);
867 case X86II::AddRegFrm:
868 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
871 case X86II::MRMDestReg:
872 EmitByte(BaseOpcode, CurByte, OS);
873 EmitRegModRMByte(MI.getOperand(CurOp),
874 GetX86RegNum(MI.getOperand(CurOp+1)), CurByte, OS);
878 case X86II::MRMDestMem:
879 EmitByte(BaseOpcode, CurByte, OS);
880 SrcRegNum = CurOp + X86::AddrNumOperands;
882 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
885 EmitMemModRMByte(MI, CurOp,
886 GetX86RegNum(MI.getOperand(SrcRegNum)),
887 TSFlags, CurByte, OS, Fixups);
888 CurOp = SrcRegNum + 1;
891 case X86II::MRMSrcReg:
892 EmitByte(BaseOpcode, CurByte, OS);
893 SrcRegNum = CurOp + 1;
895 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
898 EmitRegModRMByte(MI.getOperand(SrcRegNum),
899 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
900 CurOp = SrcRegNum + 1;
903 case X86II::MRMSrcMem: {
904 int AddrOperands = X86::AddrNumOperands;
905 unsigned FirstMemOp = CurOp+1;
908 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
911 EmitByte(BaseOpcode, CurByte, OS);
913 EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
914 TSFlags, CurByte, OS, Fixups);
915 CurOp += AddrOperands + 1;
919 case X86II::MRM0r: case X86II::MRM1r:
920 case X86II::MRM2r: case X86II::MRM3r:
921 case X86II::MRM4r: case X86II::MRM5r:
922 case X86II::MRM6r: case X86II::MRM7r:
923 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
925 EmitByte(BaseOpcode, CurByte, OS);
926 EmitRegModRMByte(MI.getOperand(CurOp++),
927 (TSFlags & X86II::FormMask)-X86II::MRM0r,
930 case X86II::MRM0m: case X86II::MRM1m:
931 case X86II::MRM2m: case X86II::MRM3m:
932 case X86II::MRM4m: case X86II::MRM5m:
933 case X86II::MRM6m: case X86II::MRM7m:
934 EmitByte(BaseOpcode, CurByte, OS);
935 EmitMemModRMByte(MI, CurOp, (TSFlags & X86II::FormMask)-X86II::MRM0m,
936 TSFlags, CurByte, OS, Fixups);
937 CurOp += X86::AddrNumOperands;
940 EmitByte(BaseOpcode, CurByte, OS);
941 EmitByte(0xC1, CurByte, OS);
944 EmitByte(BaseOpcode, CurByte, OS);
945 EmitByte(0xC2, CurByte, OS);
948 EmitByte(BaseOpcode, CurByte, OS);
949 EmitByte(0xC3, CurByte, OS);
952 EmitByte(BaseOpcode, CurByte, OS);
953 EmitByte(0xC4, CurByte, OS);
956 EmitByte(BaseOpcode, CurByte, OS);
957 EmitByte(0xC8, CurByte, OS);
960 EmitByte(BaseOpcode, CurByte, OS);
961 EmitByte(0xC9, CurByte, OS);
964 EmitByte(BaseOpcode, CurByte, OS);
965 EmitByte(0xE8, CurByte, OS);
968 EmitByte(BaseOpcode, CurByte, OS);
969 EmitByte(0xF0, CurByte, OS);
972 EmitByte(BaseOpcode, CurByte, OS);
973 EmitByte(0xF8, CurByte, OS);
976 EmitByte(BaseOpcode, CurByte, OS);
977 EmitByte(0xF9, CurByte, OS);
981 // If there is a remaining operand, it must be a trailing immediate. Emit it
982 // according to the right size for the instruction.
983 if (CurOp != NumOps) {
984 // The last source register of a 4 operand instruction in AVX is encoded
985 // in bits[7:4] of a immediate byte, and bits[3:0] are ignored.
986 if ((TSFlags >> 32) & X86II::VEX_I8IMM) {
987 const MCOperand &MO = MI.getOperand(CurOp++);
989 X86InstrInfo::isX86_64ExtendedReg(MO.getReg());
990 unsigned RegNum = (IsExtReg ? (1 << 7) : 0);
991 RegNum |= GetX86RegNum(MO) << 4;
992 EmitImmediate(MCOperand::CreateImm(RegNum), 1, FK_Data_1, CurByte, OS,
996 if (MI.getOpcode() == X86::MOV64ri32 || MI.getOpcode() == X86::MOV64mi32)
997 FixupKind = X86::reloc_signed_4byte;
999 FixupKind = getImmFixupKind(TSFlags);
1000 EmitImmediate(MI.getOperand(CurOp++),
1001 X86II::getSizeOfImm(TSFlags), MCFixupKind(FixupKind),
1002 CurByte, OS, Fixups);
1006 if ((TSFlags >> 32) & X86II::Has3DNow0F0FOpcode)
1007 EmitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS);
1012 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
1013 errs() << "Cannot encode all operands of: ";