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_riprel_4byte", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel },
49 { "reloc_riprel_4byte_movq_load", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel }
52 if (Kind < FirstTargetFixupKind)
53 return MCCodeEmitter::getFixupKindInfo(Kind);
55 assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
57 return Infos[Kind - FirstTargetFixupKind];
60 static unsigned GetX86RegNum(const MCOperand &MO) {
61 return X86RegisterInfo::getX86RegNum(MO.getReg());
64 void EmitByte(unsigned char C, unsigned &CurByte, raw_ostream &OS) const {
69 void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
70 raw_ostream &OS) const {
71 // Output the constant in little endian byte order.
72 for (unsigned i = 0; i != Size; ++i) {
73 EmitByte(Val & 255, CurByte, OS);
78 void EmitImmediate(const MCOperand &Disp,
79 unsigned ImmSize, MCFixupKind FixupKind,
80 unsigned &CurByte, raw_ostream &OS,
81 SmallVectorImpl<MCFixup> &Fixups,
82 int ImmOffset = 0) const;
84 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
86 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
87 return RM | (RegOpcode << 3) | (Mod << 6);
90 void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
91 unsigned &CurByte, raw_ostream &OS) const {
92 EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS);
95 void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base,
96 unsigned &CurByte, raw_ostream &OS) const {
97 // SIB byte is in the same format as the ModRMByte.
98 EmitByte(ModRMByte(SS, Index, Base), CurByte, OS);
102 void EmitSegmentOverridePrefix(const MCOperand &Op, unsigned TSFlags,
103 unsigned &CurByte, raw_ostream &OS) const;
105 void EmitMemModRMByte(const MCInst &MI, unsigned Op,
106 unsigned RegOpcodeField,
107 uint64_t TSFlags, unsigned &CurByte, raw_ostream &OS,
108 SmallVectorImpl<MCFixup> &Fixups) const;
110 void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
111 SmallVectorImpl<MCFixup> &Fixups) const;
113 void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
114 const MCInst &MI, const TargetInstrDesc &Desc,
115 raw_ostream &OS) const;
117 void EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
118 const MCInst &MI, const TargetInstrDesc &Desc,
119 raw_ostream &OS) const;
122 } // end anonymous namespace
125 MCCodeEmitter *llvm::createX86_32MCCodeEmitter(const Target &,
128 return new X86MCCodeEmitter(TM, Ctx, false);
131 MCCodeEmitter *llvm::createX86_64MCCodeEmitter(const Target &,
134 return new X86MCCodeEmitter(TM, Ctx, true);
138 /// isDisp8 - Return true if this signed displacement fits in a 8-bit
139 /// sign-extended field.
140 static bool isDisp8(int Value) {
141 return Value == (signed char)Value;
144 /// getImmFixupKind - Return the appropriate fixup kind to use for an immediate
145 /// in an instruction with the specified TSFlags.
146 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
147 unsigned Size = X86II::getSizeOfImm(TSFlags);
148 bool isPCRel = X86II::isImmPCRel(TSFlags);
151 default: assert(0 && "Unknown immediate size");
152 case 1: return isPCRel ? MCFixupKind(X86::reloc_pcrel_1byte) : FK_Data_1;
153 case 4: return isPCRel ? MCFixupKind(X86::reloc_pcrel_4byte) : FK_Data_4;
154 case 2: assert(!isPCRel); return FK_Data_2;
155 case 8: assert(!isPCRel); return FK_Data_8;
160 void X86MCCodeEmitter::
161 EmitImmediate(const MCOperand &DispOp, unsigned Size, MCFixupKind FixupKind,
162 unsigned &CurByte, raw_ostream &OS,
163 SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
164 // If this is a simple integer displacement that doesn't require a relocation,
166 if (DispOp.isImm()) {
167 // FIXME: is this right for pc-rel encoding?? Probably need to emit this as
169 EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
173 // If we have an immoffset, add it to the expression.
174 const MCExpr *Expr = DispOp.getExpr();
176 // If the fixup is pc-relative, we need to bias the value to be relative to
177 // the start of the field, not the end of the field.
178 if (FixupKind == MCFixupKind(X86::reloc_pcrel_4byte) ||
179 FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
180 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load))
182 if (FixupKind == MCFixupKind(X86::reloc_pcrel_1byte))
186 Expr = MCBinaryExpr::CreateAdd(Expr, MCConstantExpr::Create(ImmOffset, Ctx),
189 // Emit a symbolic constant as a fixup and 4 zeros.
190 Fixups.push_back(MCFixup::Create(CurByte, Expr, FixupKind));
191 EmitConstant(0, Size, CurByte, OS);
194 void X86MCCodeEmitter::EmitSegmentOverridePrefix(const MCOperand &Op,
197 raw_ostream &OS) const {
198 // If no segment register is present, we don't need anything.
199 if (Op.getReg() == 0)
202 // Check if we need an override.
203 switch (Op.getReg()) {
204 case X86::CS: EmitByte(0x2E, CurByte, OS); return;
205 case X86::SS: EmitByte(0x36, CurByte, OS); return;
206 case X86::DS: EmitByte(0x3E, CurByte, OS); return;
207 case X86::ES: EmitByte(0x26, CurByte, OS); return;
208 case X86::FS: EmitByte(0x64, CurByte, OS); return;
209 case X86::GS: EmitByte(0x65, CurByte, OS); return;
212 assert(0 && "Invalid segment register!");
215 void X86MCCodeEmitter::EmitMemModRMByte(const MCInst &MI, unsigned Op,
216 unsigned RegOpcodeField,
217 uint64_t TSFlags, unsigned &CurByte,
219 SmallVectorImpl<MCFixup> &Fixups) const{
220 const MCOperand &Disp = MI.getOperand(Op+3);
221 const MCOperand &Base = MI.getOperand(Op);
222 const MCOperand &Scale = MI.getOperand(Op+1);
223 const MCOperand &IndexReg = MI.getOperand(Op+2);
224 unsigned BaseReg = Base.getReg();
226 // Handle %rip relative addressing.
227 if (BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
228 assert(Is64BitMode && "Rip-relative addressing requires 64-bit mode");
229 assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
230 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
232 unsigned FixupKind = X86::reloc_riprel_4byte;
234 // movq loads are handled with a special relocation form which allows the
235 // linker to eliminate some loads for GOT references which end up in the
236 // same linkage unit.
237 if (MI.getOpcode() == X86::MOV64rm ||
238 MI.getOpcode() == X86::MOV64rm_TC)
239 FixupKind = X86::reloc_riprel_4byte_movq_load;
241 // rip-relative addressing is actually relative to the *next* instruction.
242 // Since an immediate can follow the mod/rm byte for an instruction, this
243 // means that we need to bias the immediate field of the instruction with
244 // the size of the immediate field. If we have this case, add it into the
245 // expression to emit.
246 int ImmSize = X86II::hasImm(TSFlags) ? X86II::getSizeOfImm(TSFlags) : 0;
248 EmitImmediate(Disp, 4, MCFixupKind(FixupKind),
249 CurByte, OS, Fixups, -ImmSize);
253 unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
255 // Determine whether a SIB byte is needed.
256 // If no BaseReg, issue a RIP relative instruction only if the MCE can
257 // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
258 // 2-7) and absolute references.
260 if (// The SIB byte must be used if there is an index register.
261 IndexReg.getReg() == 0 &&
262 // The SIB byte must be used if the base is ESP/RSP/R12, all of which
263 // encode to an R/M value of 4, which indicates that a SIB byte is
265 BaseRegNo != N86::ESP &&
266 // If there is no base register and we're in 64-bit mode, we need a SIB
267 // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
268 (!Is64BitMode || BaseReg != 0)) {
270 if (BaseReg == 0) { // [disp32] in X86-32 mode
271 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
272 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
276 // If the base is not EBP/ESP and there is no displacement, use simple
277 // indirect register encoding, this handles addresses like [EAX]. The
278 // encoding for [EBP] with no displacement means [disp32] so we handle it
279 // by emitting a displacement of 0 below.
280 if (Disp.isImm() && Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
281 EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
285 // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
286 if (Disp.isImm() && isDisp8(Disp.getImm())) {
287 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
288 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
292 // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
293 EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
294 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
298 // We need a SIB byte, so start by outputting the ModR/M byte first
299 assert(IndexReg.getReg() != X86::ESP &&
300 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
302 bool ForceDisp32 = false;
303 bool ForceDisp8 = false;
305 // If there is no base register, we emit the special case SIB byte with
306 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
307 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
309 } else if (!Disp.isImm()) {
310 // Emit the normal disp32 encoding.
311 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
313 } else if (Disp.getImm() == 0 &&
314 // Base reg can't be anything that ends up with '5' as the base
315 // reg, it is the magic [*] nomenclature that indicates no base.
316 BaseRegNo != N86::EBP) {
317 // Emit no displacement ModR/M byte
318 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
319 } else if (isDisp8(Disp.getImm())) {
320 // Emit the disp8 encoding.
321 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
322 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
324 // Emit the normal disp32 encoding.
325 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
328 // Calculate what the SS field value should be...
329 static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 };
330 unsigned SS = SSTable[Scale.getImm()];
333 // Handle the SIB byte for the case where there is no base, see Intel
334 // Manual 2A, table 2-7. The displacement has already been output.
336 if (IndexReg.getReg())
337 IndexRegNo = GetX86RegNum(IndexReg);
338 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
340 EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
343 if (IndexReg.getReg())
344 IndexRegNo = GetX86RegNum(IndexReg);
346 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
347 EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
350 // Do we need to output a displacement?
352 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
353 else if (ForceDisp32 || Disp.getImm() != 0)
354 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
357 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
359 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
360 const MCInst &MI, const TargetInstrDesc &Desc,
361 raw_ostream &OS) const {
363 // Pseudo instructions never have a VEX prefix.
364 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
367 bool HasVEX_4V = false;
368 if ((TSFlags >> 32) & X86II::VEX_4V)
371 // VEX_R: opcode externsion equivalent to REX.R in
372 // 1's complement (inverted) form
374 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
375 // 0: Same as REX_R=1 (64 bit mode only)
377 unsigned char VEX_R = 0x1;
379 // VEX_X: equivalent to REX.X, only used when a
380 // register is used for index in SIB Byte.
382 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
383 // 0: Same as REX.X=1 (64-bit mode only)
384 unsigned char VEX_X = 0x1;
388 // 1: Same as REX_B=0 (ignored in 32-bit mode)
389 // 0: Same as REX_B=1 (64 bit mode only)
391 unsigned char VEX_B = 0x1;
393 // VEX_W: opcode specific (use like REX.W, or used for
394 // opcode extension, or ignored, depending on the opcode byte)
395 unsigned char VEX_W = 0;
397 // VEX_5M (VEX m-mmmmm field):
399 // 0b00000: Reserved for future use
400 // 0b00001: implied 0F leading opcode
401 // 0b00010: implied 0F 38 leading opcode bytes
402 // 0b00011: implied 0F 3A leading opcode bytes
403 // 0b00100-0b11111: Reserved for future use
405 unsigned char VEX_5M = 0x1;
407 // VEX_4V (VEX vvvv field): a register specifier
408 // (in 1's complement form) or 1111 if unused.
409 unsigned char VEX_4V = 0xf;
411 // VEX_L (Vector Length):
413 // 0: scalar or 128-bit vector
416 unsigned char VEX_L = 0;
418 // VEX_PP: opcode extension providing equivalent
419 // functionality of a SIMD prefix
426 unsigned char VEX_PP = 0;
428 // Encode the operand size opcode prefix as needed.
429 if (TSFlags & X86II::OpSize)
432 switch (TSFlags & X86II::Op0Mask) {
433 default: assert(0 && "Invalid prefix!");
434 case X86II::T8: // 0F 38
437 case X86II::TA: // 0F 3A
440 case X86II::TF: // F2 0F 38
444 case X86II::XS: // F3 0F
447 case X86II::XD: // F2 0F
450 case X86II::TB: // Bypass: Not used by VEX
455 unsigned NumOps = MI.getNumOperands();
458 if ((TSFlags & X86II::FormMask) == X86II::MRMDestMem)
459 NumOps = CurOp = X86AddrNumOperands;
461 switch (TSFlags & X86II::FormMask) {
462 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
463 case X86II::MRMSrcMem:
464 case X86II::MRMDestMem:
465 case X86II::MRMSrcReg:
466 if (MI.getOperand(CurOp).isReg() &&
467 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
470 // If the memory destination has been checked first,
471 // go back to the first operand
472 CurOp = (CurOp == NumOps) ? 0 : CurOp+1;
474 // On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the
475 // range 0-7 and the difference between the 2 groups is given by the
476 // REX prefix. In the VEX prefix, registers are seen sequencially
477 // from 0-15 and encoded in 1's complement form, example:
479 // ModRM field => XMM9 => 1
480 // VEX.VVVV => XMM9 => ~9
482 // See table 4-35 of Intel AVX Programming Reference for details.
484 unsigned SrcReg = MI.getOperand(CurOp).getReg();
485 unsigned SrcRegNum = GetX86RegNum(MI.getOperand(1));
486 if (SrcReg >= X86::XMM8 && SrcReg <= X86::XMM15)
489 // The registers represented through VEX_VVVV should
490 // be encoded in 1's complement form.
491 VEX_4V = (~SrcRegNum) & 0xf;
496 for (; CurOp != NumOps; ++CurOp) {
497 const MCOperand &MO = MI.getOperand(CurOp);
498 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
500 if (!VEX_B && MO.isReg() &&
501 ((TSFlags & X86II::FormMask) == X86II::MRMSrcMem) &&
502 X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
507 assert(0 && "Not implemented!");
510 // VEX opcode prefix can have 2 or 3 bytes
513 // +-----+ +--------------+ +-------------------+
514 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
515 // +-----+ +--------------+ +-------------------+
517 // +-----+ +-------------------+
518 // | C5h | | R | vvvv | L | pp |
519 // +-----+ +-------------------+
521 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
523 if (VEX_B && VEX_X) { // 2 byte VEX prefix
524 EmitByte(0xC5, CurByte, OS);
525 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
530 EmitByte(0xC4, CurByte, OS);
531 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_5M, CurByte, OS);
532 EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
535 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
536 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
537 /// size, and 3) use of X86-64 extended registers.
538 static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
539 const TargetInstrDesc &Desc) {
540 // Pseudo instructions never have a rex byte.
541 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
545 if (TSFlags & X86II::REX_W)
546 REX |= 1 << 3; // set REX.W
548 if (MI.getNumOperands() == 0) return REX;
550 unsigned NumOps = MI.getNumOperands();
551 // FIXME: MCInst should explicitize the two-addrness.
552 bool isTwoAddr = NumOps > 1 &&
553 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
555 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
556 unsigned i = isTwoAddr ? 1 : 0;
557 for (; i != NumOps; ++i) {
558 const MCOperand &MO = MI.getOperand(i);
559 if (!MO.isReg()) continue;
560 unsigned Reg = MO.getReg();
561 if (!X86InstrInfo::isX86_64NonExtLowByteReg(Reg)) continue;
562 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
563 // that returns non-zero.
564 REX |= 0x40; // REX fixed encoding prefix
568 switch (TSFlags & X86II::FormMask) {
569 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
570 case X86II::MRMSrcReg:
571 if (MI.getOperand(0).isReg() &&
572 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
573 REX |= 1 << 2; // set REX.R
574 i = isTwoAddr ? 2 : 1;
575 for (; i != NumOps; ++i) {
576 const MCOperand &MO = MI.getOperand(i);
577 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
578 REX |= 1 << 0; // set REX.B
581 case X86II::MRMSrcMem: {
582 if (MI.getOperand(0).isReg() &&
583 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
584 REX |= 1 << 2; // set REX.R
586 i = isTwoAddr ? 2 : 1;
587 for (; i != NumOps; ++i) {
588 const MCOperand &MO = MI.getOperand(i);
590 if (X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
591 REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1)
597 case X86II::MRM0m: case X86II::MRM1m:
598 case X86II::MRM2m: case X86II::MRM3m:
599 case X86II::MRM4m: case X86II::MRM5m:
600 case X86II::MRM6m: case X86II::MRM7m:
601 case X86II::MRMDestMem: {
602 unsigned e = (isTwoAddr ? X86AddrNumOperands+1 : X86AddrNumOperands);
603 i = isTwoAddr ? 1 : 0;
604 if (NumOps > e && MI.getOperand(e).isReg() &&
605 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(e).getReg()))
606 REX |= 1 << 2; // set REX.R
608 for (; i != e; ++i) {
609 const MCOperand &MO = MI.getOperand(i);
611 if (X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
612 REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1)
619 if (MI.getOperand(0).isReg() &&
620 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
621 REX |= 1 << 0; // set REX.B
622 i = isTwoAddr ? 2 : 1;
623 for (unsigned e = NumOps; i != e; ++i) {
624 const MCOperand &MO = MI.getOperand(i);
625 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
626 REX |= 1 << 2; // set REX.R
633 /// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode.
634 void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
635 const MCInst &MI, const TargetInstrDesc &Desc,
636 raw_ostream &OS) const {
638 // Emit the lock opcode prefix as needed.
639 if (TSFlags & X86II::LOCK)
640 EmitByte(0xF0, CurByte, OS);
642 // Emit segment override opcode prefix as needed.
643 switch (TSFlags & X86II::SegOvrMask) {
644 default: assert(0 && "Invalid segment!");
645 case 0: break; // No segment override!
647 EmitByte(0x64, CurByte, OS);
650 EmitByte(0x65, CurByte, OS);
654 // Emit the repeat opcode prefix as needed.
655 if ((TSFlags & X86II::Op0Mask) == X86II::REP)
656 EmitByte(0xF3, CurByte, OS);
658 // Emit the operand size opcode prefix as needed.
659 if (TSFlags & X86II::OpSize)
660 EmitByte(0x66, CurByte, OS);
662 // Emit the address size opcode prefix as needed.
663 if (TSFlags & X86II::AdSize)
664 EmitByte(0x67, CurByte, OS);
666 bool Need0FPrefix = false;
667 switch (TSFlags & X86II::Op0Mask) {
668 default: assert(0 && "Invalid prefix!");
669 case 0: break; // No prefix!
670 case X86II::REP: break; // already handled.
671 case X86II::TB: // Two-byte opcode prefix
672 case X86II::T8: // 0F 38
673 case X86II::TA: // 0F 3A
676 case X86II::TF: // F2 0F 38
677 EmitByte(0xF2, CurByte, OS);
680 case X86II::XS: // F3 0F
681 EmitByte(0xF3, CurByte, OS);
684 case X86II::XD: // F2 0F
685 EmitByte(0xF2, CurByte, OS);
688 case X86II::D8: EmitByte(0xD8, CurByte, OS); break;
689 case X86II::D9: EmitByte(0xD9, CurByte, OS); break;
690 case X86II::DA: EmitByte(0xDA, CurByte, OS); break;
691 case X86II::DB: EmitByte(0xDB, CurByte, OS); break;
692 case X86II::DC: EmitByte(0xDC, CurByte, OS); break;
693 case X86II::DD: EmitByte(0xDD, CurByte, OS); break;
694 case X86II::DE: EmitByte(0xDE, CurByte, OS); break;
695 case X86II::DF: EmitByte(0xDF, CurByte, OS); break;
698 // Handle REX prefix.
699 // FIXME: Can this come before F2 etc to simplify emission?
701 if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc))
702 EmitByte(0x40 | REX, CurByte, OS);
705 // 0x0F escape code must be emitted just before the opcode.
707 EmitByte(0x0F, CurByte, OS);
709 // FIXME: Pull this up into previous switch if REX can be moved earlier.
710 switch (TSFlags & X86II::Op0Mask) {
711 case X86II::TF: // F2 0F 38
712 case X86II::T8: // 0F 38
713 EmitByte(0x38, CurByte, OS);
715 case X86II::TA: // 0F 3A
716 EmitByte(0x3A, CurByte, OS);
721 void X86MCCodeEmitter::
722 EncodeInstruction(const MCInst &MI, raw_ostream &OS,
723 SmallVectorImpl<MCFixup> &Fixups) const {
724 unsigned Opcode = MI.getOpcode();
725 const TargetInstrDesc &Desc = TII.get(Opcode);
726 uint64_t TSFlags = Desc.TSFlags;
728 // Keep track of the current byte being emitted.
729 unsigned CurByte = 0;
731 // Is this instruction encoded using the AVX VEX prefix?
732 bool HasVEXPrefix = false;
734 // It uses the VEX.VVVV field?
735 bool HasVEX_4V = false;
737 if ((TSFlags >> 32) & X86II::VEX)
739 if ((TSFlags >> 32) & X86II::VEX_4V)
742 // FIXME: We should emit the prefixes in exactly the same order as GAS does,
743 // in order to provide diffability.
746 EmitOpcodePrefix(TSFlags, CurByte, MI, Desc, OS);
748 EmitVEXOpcodePrefix(TSFlags, CurByte, MI, Desc, OS);
750 // If this is a two-address instruction, skip one of the register operands.
751 unsigned NumOps = Desc.getNumOperands();
753 if (NumOps > 1 && Desc.getOperandConstraint(1, TOI::TIED_TO) != -1)
755 else if (NumOps > 2 && Desc.getOperandConstraint(NumOps-1, TOI::TIED_TO)== 0)
756 // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32
759 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
760 unsigned SrcRegNum = 0;
761 switch (TSFlags & X86II::FormMask) {
762 case X86II::MRMInitReg:
763 assert(0 && "FIXME: Remove this form when the JIT moves to MCCodeEmitter!");
764 default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
765 assert(0 && "Unknown FormMask value in X86MCCodeEmitter!");
766 case X86II::Pseudo: return; // Pseudo instructions encode to nothing.
768 EmitByte(BaseOpcode, CurByte, OS);
771 case X86II::AddRegFrm:
772 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
775 case X86II::MRMDestReg:
776 EmitByte(BaseOpcode, CurByte, OS);
777 EmitRegModRMByte(MI.getOperand(CurOp),
778 GetX86RegNum(MI.getOperand(CurOp+1)), CurByte, OS);
782 case X86II::MRMDestMem:
783 EmitSegmentOverridePrefix(MI.getOperand(CurOp + 4), TSFlags, CurByte, OS);
784 EmitByte(BaseOpcode, CurByte, OS);
785 EmitMemModRMByte(MI, CurOp,
786 GetX86RegNum(MI.getOperand(CurOp + X86AddrNumOperands)),
787 TSFlags, CurByte, OS, Fixups);
788 CurOp += X86AddrNumOperands + 1;
791 case X86II::MRMSrcReg:
792 EmitByte(BaseOpcode, CurByte, OS);
793 SrcRegNum = CurOp + 1;
795 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
798 EmitRegModRMByte(MI.getOperand(SrcRegNum),
799 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
800 CurOp = SrcRegNum + 1;
803 case X86II::MRMSrcMem: {
804 int AddrOperands = X86AddrNumOperands;
805 unsigned FirstMemOp = CurOp+1;
808 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
811 // FIXME: Maybe lea should have its own form? This is a horrible hack.
812 if (Opcode == X86::LEA64r || Opcode == X86::LEA64_32r ||
813 Opcode == X86::LEA16r || Opcode == X86::LEA32r)
814 --AddrOperands; // No segment register
816 EmitSegmentOverridePrefix(MI.getOperand(FirstMemOp+4),
817 TSFlags, CurByte, OS);
819 EmitByte(BaseOpcode, CurByte, OS);
822 EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
823 TSFlags, CurByte, OS, Fixups);
824 CurOp += AddrOperands + 1;
828 case X86II::MRM0r: case X86II::MRM1r:
829 case X86II::MRM2r: case X86II::MRM3r:
830 case X86II::MRM4r: case X86II::MRM5r:
831 case X86II::MRM6r: case X86II::MRM7r:
832 EmitByte(BaseOpcode, CurByte, OS);
833 EmitRegModRMByte(MI.getOperand(CurOp++),
834 (TSFlags & X86II::FormMask)-X86II::MRM0r,
837 case X86II::MRM0m: case X86II::MRM1m:
838 case X86II::MRM2m: case X86II::MRM3m:
839 case X86II::MRM4m: case X86II::MRM5m:
840 case X86II::MRM6m: case X86II::MRM7m:
841 EmitSegmentOverridePrefix(MI.getOperand(CurOp+4), TSFlags, CurByte, OS);
842 EmitByte(BaseOpcode, CurByte, OS);
843 EmitMemModRMByte(MI, CurOp, (TSFlags & X86II::FormMask)-X86II::MRM0m,
844 TSFlags, CurByte, OS, Fixups);
845 CurOp += X86AddrNumOperands;
848 EmitByte(BaseOpcode, CurByte, OS);
849 EmitByte(0xC1, CurByte, OS);
852 EmitByte(BaseOpcode, CurByte, OS);
853 EmitByte(0xC2, CurByte, OS);
856 EmitByte(BaseOpcode, CurByte, OS);
857 EmitByte(0xC3, CurByte, OS);
860 EmitByte(BaseOpcode, CurByte, OS);
861 EmitByte(0xC4, CurByte, OS);
864 EmitByte(BaseOpcode, CurByte, OS);
865 EmitByte(0xC8, CurByte, OS);
868 EmitByte(BaseOpcode, CurByte, OS);
869 EmitByte(0xC9, CurByte, OS);
872 EmitByte(BaseOpcode, CurByte, OS);
873 EmitByte(0xE8, CurByte, OS);
876 EmitByte(BaseOpcode, CurByte, OS);
877 EmitByte(0xF0, CurByte, OS);
880 EmitByte(BaseOpcode, CurByte, OS);
881 EmitByte(0xF8, CurByte, OS);
884 EmitByte(BaseOpcode, CurByte, OS);
885 EmitByte(0xF9, CurByte, OS);
889 // If there is a remaining operand, it must be a trailing immediate. Emit it
890 // according to the right size for the instruction.
892 EmitImmediate(MI.getOperand(CurOp++),
893 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
894 CurByte, OS, Fixups);
898 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
899 errs() << "Cannot encode all operands of: ";