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 }
53 if (Kind < FirstTargetFixupKind)
54 return MCCodeEmitter::getFixupKindInfo(Kind);
56 assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
58 return Infos[Kind - FirstTargetFixupKind];
61 static unsigned GetX86RegNum(const MCOperand &MO) {
62 return X86RegisterInfo::getX86RegNum(MO.getReg());
65 // On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the range
66 // 0-7 and the difference between the 2 groups is given by the REX prefix.
67 // In the VEX prefix, registers are seen sequencially from 0-15 and encoded
68 // in 1's complement form, example:
70 // ModRM field => XMM9 => 1
71 // VEX.VVVV => XMM9 => ~9
73 // See table 4-35 of Intel AVX Programming Reference for details.
74 static unsigned char getVEXRegisterEncoding(const MCInst &MI,
76 unsigned SrcReg = MI.getOperand(OpNum).getReg();
77 unsigned SrcRegNum = GetX86RegNum(MI.getOperand(OpNum));
78 if ((SrcReg >= X86::XMM8 && SrcReg <= X86::XMM15) ||
79 (SrcReg >= X86::YMM8 && SrcReg <= X86::YMM15))
82 // The registers represented through VEX_VVVV should
83 // be encoded in 1's complement form.
84 return (~SrcRegNum) & 0xf;
87 void EmitByte(unsigned char C, unsigned &CurByte, raw_ostream &OS) const {
92 void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
93 raw_ostream &OS) const {
94 // Output the constant in little endian byte order.
95 for (unsigned i = 0; i != Size; ++i) {
96 EmitByte(Val & 255, CurByte, OS);
101 void EmitImmediate(const MCOperand &Disp,
102 unsigned ImmSize, MCFixupKind FixupKind,
103 unsigned &CurByte, raw_ostream &OS,
104 SmallVectorImpl<MCFixup> &Fixups,
105 int ImmOffset = 0) const;
107 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
109 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
110 return RM | (RegOpcode << 3) | (Mod << 6);
113 void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
114 unsigned &CurByte, raw_ostream &OS) const {
115 EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS);
118 void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base,
119 unsigned &CurByte, raw_ostream &OS) const {
120 // SIB byte is in the same format as the ModRMByte.
121 EmitByte(ModRMByte(SS, Index, Base), CurByte, OS);
125 void EmitMemModRMByte(const MCInst &MI, unsigned Op,
126 unsigned RegOpcodeField,
127 uint64_t TSFlags, unsigned &CurByte, raw_ostream &OS,
128 SmallVectorImpl<MCFixup> &Fixups) const;
130 void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
131 SmallVectorImpl<MCFixup> &Fixups) const;
133 void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
134 const MCInst &MI, const TargetInstrDesc &Desc,
135 raw_ostream &OS) const;
137 void EmitSegmentOverridePrefix(uint64_t TSFlags, unsigned &CurByte,
138 int MemOperand, const MCInst &MI,
139 raw_ostream &OS) const;
141 void EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
142 const MCInst &MI, const TargetInstrDesc &Desc,
143 raw_ostream &OS) const;
146 } // end anonymous namespace
149 MCCodeEmitter *llvm::createX86_32MCCodeEmitter(const Target &,
152 return new X86MCCodeEmitter(TM, Ctx, false);
155 MCCodeEmitter *llvm::createX86_64MCCodeEmitter(const Target &,
158 return new X86MCCodeEmitter(TM, Ctx, true);
161 /// isDisp8 - Return true if this signed displacement fits in a 8-bit
162 /// sign-extended field.
163 static bool isDisp8(int Value) {
164 return Value == (signed char)Value;
167 /// getImmFixupKind - Return the appropriate fixup kind to use for an immediate
168 /// in an instruction with the specified TSFlags.
169 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
170 unsigned Size = X86II::getSizeOfImm(TSFlags);
171 bool isPCRel = X86II::isImmPCRel(TSFlags);
174 default: assert(0 && "Unknown immediate size");
175 case 1: return isPCRel ? MCFixupKind(X86::reloc_pcrel_1byte) : FK_Data_1;
176 case 2: return isPCRel ? MCFixupKind(X86::reloc_pcrel_2byte) : FK_Data_2;
177 case 4: return isPCRel ? MCFixupKind(X86::reloc_pcrel_4byte) : FK_Data_4;
178 case 8: assert(!isPCRel); return FK_Data_8;
183 void X86MCCodeEmitter::
184 EmitImmediate(const MCOperand &DispOp, unsigned Size, MCFixupKind FixupKind,
185 unsigned &CurByte, raw_ostream &OS,
186 SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
187 // If this is a simple integer displacement that doesn't require a relocation,
189 if (DispOp.isImm()) {
190 // FIXME: is this right for pc-rel encoding?? Probably need to emit this as
192 EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
196 // If we have an immoffset, add it to the expression.
197 const MCExpr *Expr = DispOp.getExpr();
199 // If the fixup is pc-relative, we need to bias the value to be relative to
200 // the start of the field, not the end of the field.
201 if (FixupKind == MCFixupKind(X86::reloc_pcrel_4byte) ||
202 FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
203 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load))
205 if (FixupKind == MCFixupKind(X86::reloc_pcrel_2byte))
207 if (FixupKind == MCFixupKind(X86::reloc_pcrel_1byte))
211 Expr = MCBinaryExpr::CreateAdd(Expr, MCConstantExpr::Create(ImmOffset, Ctx),
214 // Emit a symbolic constant as a fixup and 4 zeros.
215 Fixups.push_back(MCFixup::Create(CurByte, Expr, FixupKind));
216 EmitConstant(0, Size, CurByte, OS);
219 void X86MCCodeEmitter::EmitMemModRMByte(const MCInst &MI, unsigned Op,
220 unsigned RegOpcodeField,
221 uint64_t TSFlags, unsigned &CurByte,
223 SmallVectorImpl<MCFixup> &Fixups) const{
224 const MCOperand &Disp = MI.getOperand(Op+3);
225 const MCOperand &Base = MI.getOperand(Op);
226 const MCOperand &Scale = MI.getOperand(Op+1);
227 const MCOperand &IndexReg = MI.getOperand(Op+2);
228 unsigned BaseReg = Base.getReg();
230 // Handle %rip relative addressing.
231 if (BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
232 assert(Is64BitMode && "Rip-relative addressing requires 64-bit mode");
233 assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
234 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
236 unsigned FixupKind = X86::reloc_riprel_4byte;
238 // movq loads are handled with a special relocation form which allows the
239 // linker to eliminate some loads for GOT references which end up in the
240 // same linkage unit.
241 if (MI.getOpcode() == X86::MOV64rm ||
242 MI.getOpcode() == X86::MOV64rm_TC)
243 FixupKind = X86::reloc_riprel_4byte_movq_load;
245 // rip-relative addressing is actually relative to the *next* instruction.
246 // Since an immediate can follow the mod/rm byte for an instruction, this
247 // means that we need to bias the immediate field of the instruction with
248 // the size of the immediate field. If we have this case, add it into the
249 // expression to emit.
250 int ImmSize = X86II::hasImm(TSFlags) ? X86II::getSizeOfImm(TSFlags) : 0;
252 EmitImmediate(Disp, 4, MCFixupKind(FixupKind),
253 CurByte, OS, Fixups, -ImmSize);
257 unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
259 // Determine whether a SIB byte is needed.
260 // If no BaseReg, issue a RIP relative instruction only if the MCE can
261 // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
262 // 2-7) and absolute references.
264 if (// The SIB byte must be used if there is an index register.
265 IndexReg.getReg() == 0 &&
266 // The SIB byte must be used if the base is ESP/RSP/R12, all of which
267 // encode to an R/M value of 4, which indicates that a SIB byte is
269 BaseRegNo != N86::ESP &&
270 // If there is no base register and we're in 64-bit mode, we need a SIB
271 // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
272 (!Is64BitMode || BaseReg != 0)) {
274 if (BaseReg == 0) { // [disp32] in X86-32 mode
275 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
276 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
280 // If the base is not EBP/ESP and there is no displacement, use simple
281 // indirect register encoding, this handles addresses like [EAX]. The
282 // encoding for [EBP] with no displacement means [disp32] so we handle it
283 // by emitting a displacement of 0 below.
284 if (Disp.isImm() && Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
285 EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
289 // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
290 if (Disp.isImm() && isDisp8(Disp.getImm())) {
291 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
292 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
296 // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
297 EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
298 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
302 // We need a SIB byte, so start by outputting the ModR/M byte first
303 assert(IndexReg.getReg() != X86::ESP &&
304 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
306 bool ForceDisp32 = false;
307 bool ForceDisp8 = false;
309 // If there is no base register, we emit the special case SIB byte with
310 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
311 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
313 } else if (!Disp.isImm()) {
314 // Emit the normal disp32 encoding.
315 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
317 } else if (Disp.getImm() == 0 &&
318 // Base reg can't be anything that ends up with '5' as the base
319 // reg, it is the magic [*] nomenclature that indicates no base.
320 BaseRegNo != N86::EBP) {
321 // Emit no displacement ModR/M byte
322 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
323 } else if (isDisp8(Disp.getImm())) {
324 // Emit the disp8 encoding.
325 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
326 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
328 // Emit the normal disp32 encoding.
329 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
332 // Calculate what the SS field value should be...
333 static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 };
334 unsigned SS = SSTable[Scale.getImm()];
337 // Handle the SIB byte for the case where there is no base, see Intel
338 // Manual 2A, table 2-7. The displacement has already been output.
340 if (IndexReg.getReg())
341 IndexRegNo = GetX86RegNum(IndexReg);
342 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
344 EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
347 if (IndexReg.getReg())
348 IndexRegNo = GetX86RegNum(IndexReg);
350 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
351 EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
354 // Do we need to output a displacement?
356 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
357 else if (ForceDisp32 || Disp.getImm() != 0)
358 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
361 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
363 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
364 int MemOperand, const MCInst &MI,
365 const TargetInstrDesc &Desc,
366 raw_ostream &OS) const {
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 if ((TSFlags >> 32) & X86II::VEX_W)
435 if ((TSFlags >> 32) & X86II::VEX_L)
438 switch (TSFlags & X86II::Op0Mask) {
439 default: assert(0 && "Invalid prefix!");
440 case X86II::T8: // 0F 38
443 case X86II::TA: // 0F 3A
446 case X86II::TF: // F2 0F 38
450 case X86II::XS: // F3 0F
453 case X86II::XD: // F2 0F
456 case X86II::TB: // Bypass: Not used by VEX
461 // Set the vector length to 256-bit if YMM0-YMM15 is used
462 for (unsigned i = 0; i != MI.getNumOperands(); ++i) {
463 if (!MI.getOperand(i).isReg())
465 unsigned SrcReg = MI.getOperand(i).getReg();
466 if (SrcReg >= X86::YMM0 && SrcReg <= X86::YMM15)
470 unsigned NumOps = MI.getNumOperands();
472 bool IsDestMem = false;
474 switch (TSFlags & X86II::FormMask) {
475 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
476 case X86II::MRMDestMem:
478 // The important info for the VEX prefix is never beyond the address
479 // registers. Don't check beyond that.
480 NumOps = CurOp = X86::AddrNumOperands;
481 case X86II::MRM0m: case X86II::MRM1m:
482 case X86II::MRM2m: case X86II::MRM3m:
483 case X86II::MRM4m: case X86II::MRM5m:
484 case X86II::MRM6m: case X86II::MRM7m:
485 case X86II::MRMSrcMem:
486 case X86II::MRMSrcReg:
487 if (MI.getNumOperands() > CurOp && MI.getOperand(CurOp).isReg() &&
488 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
493 VEX_4V = getVEXRegisterEncoding(MI, IsDestMem ? CurOp-1 : CurOp);
497 // To only check operands before the memory address ones, start
498 // the search from the begining
502 // If the last register should be encoded in the immediate field
503 // do not use any bit from VEX prefix to this register, ignore it
504 if ((TSFlags >> 32) & X86II::VEX_I8IMM)
507 for (; CurOp != NumOps; ++CurOp) {
508 const MCOperand &MO = MI.getOperand(CurOp);
509 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
511 if (!VEX_B && MO.isReg() &&
512 ((TSFlags & X86II::FormMask) == X86II::MRMSrcMem) &&
513 X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
517 default: // MRMDestReg, MRM0r-MRM7r, RawFrm
518 if (!MI.getNumOperands())
521 if (MI.getOperand(CurOp).isReg() &&
522 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
526 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
529 for (; CurOp != NumOps; ++CurOp) {
530 const MCOperand &MO = MI.getOperand(CurOp);
531 if (MO.isReg() && !HasVEX_4V &&
532 X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
538 // Emit segment override opcode prefix as needed.
539 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
541 // VEX opcode prefix can have 2 or 3 bytes
544 // +-----+ +--------------+ +-------------------+
545 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
546 // +-----+ +--------------+ +-------------------+
548 // +-----+ +-------------------+
549 // | C5h | | R | vvvv | L | pp |
550 // +-----+ +-------------------+
552 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
554 if (VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) { // 2 byte VEX prefix
555 EmitByte(0xC5, CurByte, OS);
556 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
561 EmitByte(0xC4, CurByte, OS);
562 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
563 EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
566 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
567 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
568 /// size, and 3) use of X86-64 extended registers.
569 static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
570 const TargetInstrDesc &Desc) {
572 if (TSFlags & X86II::REX_W)
573 REX |= 1 << 3; // set REX.W
575 if (MI.getNumOperands() == 0) return REX;
577 unsigned NumOps = MI.getNumOperands();
578 // FIXME: MCInst should explicitize the two-addrness.
579 bool isTwoAddr = NumOps > 1 &&
580 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
582 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
583 unsigned i = isTwoAddr ? 1 : 0;
584 for (; i != NumOps; ++i) {
585 const MCOperand &MO = MI.getOperand(i);
586 if (!MO.isReg()) continue;
587 unsigned Reg = MO.getReg();
588 if (!X86InstrInfo::isX86_64NonExtLowByteReg(Reg)) continue;
589 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
590 // that returns non-zero.
591 REX |= 0x40; // REX fixed encoding prefix
595 switch (TSFlags & X86II::FormMask) {
596 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
597 case X86II::MRMSrcReg:
598 if (MI.getOperand(0).isReg() &&
599 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
600 REX |= 1 << 2; // set REX.R
601 i = isTwoAddr ? 2 : 1;
602 for (; i != NumOps; ++i) {
603 const MCOperand &MO = MI.getOperand(i);
604 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
605 REX |= 1 << 0; // set REX.B
608 case X86II::MRMSrcMem: {
609 if (MI.getOperand(0).isReg() &&
610 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
611 REX |= 1 << 2; // set REX.R
613 i = isTwoAddr ? 2 : 1;
614 for (; i != NumOps; ++i) {
615 const MCOperand &MO = MI.getOperand(i);
617 if (X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
618 REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1)
624 case X86II::MRM0m: case X86II::MRM1m:
625 case X86II::MRM2m: case X86II::MRM3m:
626 case X86II::MRM4m: case X86II::MRM5m:
627 case X86II::MRM6m: case X86II::MRM7m:
628 case X86II::MRMDestMem: {
629 unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands);
630 i = isTwoAddr ? 1 : 0;
631 if (NumOps > e && MI.getOperand(e).isReg() &&
632 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(e).getReg()))
633 REX |= 1 << 2; // set REX.R
635 for (; i != e; ++i) {
636 const MCOperand &MO = MI.getOperand(i);
638 if (X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
639 REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1)
646 if (MI.getOperand(0).isReg() &&
647 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
648 REX |= 1 << 0; // set REX.B
649 i = isTwoAddr ? 2 : 1;
650 for (unsigned e = NumOps; i != e; ++i) {
651 const MCOperand &MO = MI.getOperand(i);
652 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
653 REX |= 1 << 2; // set REX.R
660 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
661 void X86MCCodeEmitter::EmitSegmentOverridePrefix(uint64_t TSFlags,
662 unsigned &CurByte, int MemOperand,
664 raw_ostream &OS) const {
665 switch (TSFlags & X86II::SegOvrMask) {
666 default: assert(0 && "Invalid segment!");
668 // No segment override, check for explicit one on memory operand.
669 if (MemOperand != -1) { // If the instruction has a memory operand.
670 switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
671 default: assert(0 && "Unknown segment register!");
673 case X86::CS: EmitByte(0x2E, CurByte, OS); break;
674 case X86::SS: EmitByte(0x36, CurByte, OS); break;
675 case X86::DS: EmitByte(0x3E, CurByte, OS); break;
676 case X86::ES: EmitByte(0x26, CurByte, OS); break;
677 case X86::FS: EmitByte(0x64, CurByte, OS); break;
678 case X86::GS: EmitByte(0x65, CurByte, OS); break;
683 EmitByte(0x64, CurByte, OS);
686 EmitByte(0x65, CurByte, OS);
691 /// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode.
693 /// MemOperand is the operand # of the start of a memory operand if present. If
694 /// Not present, it is -1.
695 void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
696 int MemOperand, const MCInst &MI,
697 const TargetInstrDesc &Desc,
698 raw_ostream &OS) const {
700 // Emit the lock opcode prefix as needed.
701 if (TSFlags & X86II::LOCK)
702 EmitByte(0xF0, CurByte, OS);
704 // Emit segment override opcode prefix as needed.
705 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
707 // Emit the repeat opcode prefix as needed.
708 if ((TSFlags & X86II::Op0Mask) == X86II::REP)
709 EmitByte(0xF3, CurByte, OS);
711 // Emit the operand size opcode prefix as needed.
712 if (TSFlags & X86II::OpSize)
713 EmitByte(0x66, CurByte, OS);
715 // Emit the address size opcode prefix as needed.
716 if (TSFlags & X86II::AdSize)
717 EmitByte(0x67, CurByte, OS);
719 bool Need0FPrefix = false;
720 switch (TSFlags & X86II::Op0Mask) {
721 default: assert(0 && "Invalid prefix!");
722 case 0: break; // No prefix!
723 case X86II::REP: break; // already handled.
724 case X86II::TB: // Two-byte opcode prefix
725 case X86II::T8: // 0F 38
726 case X86II::TA: // 0F 3A
729 case X86II::TF: // F2 0F 38
730 EmitByte(0xF2, CurByte, OS);
733 case X86II::XS: // F3 0F
734 EmitByte(0xF3, CurByte, OS);
737 case X86II::XD: // F2 0F
738 EmitByte(0xF2, CurByte, OS);
741 case X86II::D8: EmitByte(0xD8, CurByte, OS); break;
742 case X86II::D9: EmitByte(0xD9, CurByte, OS); break;
743 case X86II::DA: EmitByte(0xDA, CurByte, OS); break;
744 case X86II::DB: EmitByte(0xDB, CurByte, OS); break;
745 case X86II::DC: EmitByte(0xDC, CurByte, OS); break;
746 case X86II::DD: EmitByte(0xDD, CurByte, OS); break;
747 case X86II::DE: EmitByte(0xDE, CurByte, OS); break;
748 case X86II::DF: EmitByte(0xDF, CurByte, OS); break;
751 // Handle REX prefix.
752 // FIXME: Can this come before F2 etc to simplify emission?
754 if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc))
755 EmitByte(0x40 | REX, CurByte, OS);
758 // 0x0F escape code must be emitted just before the opcode.
760 EmitByte(0x0F, CurByte, OS);
762 // FIXME: Pull this up into previous switch if REX can be moved earlier.
763 switch (TSFlags & X86II::Op0Mask) {
764 case X86II::TF: // F2 0F 38
765 case X86II::T8: // 0F 38
766 EmitByte(0x38, CurByte, OS);
768 case X86II::TA: // 0F 3A
769 EmitByte(0x3A, CurByte, OS);
774 void X86MCCodeEmitter::
775 EncodeInstruction(const MCInst &MI, raw_ostream &OS,
776 SmallVectorImpl<MCFixup> &Fixups) const {
777 unsigned Opcode = MI.getOpcode();
778 const TargetInstrDesc &Desc = TII.get(Opcode);
779 uint64_t TSFlags = Desc.TSFlags;
781 // Pseudo instructions don't get encoded.
782 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
785 // If this is a two-address instruction, skip one of the register operands.
786 // FIXME: This should be handled during MCInst lowering.
787 unsigned NumOps = Desc.getNumOperands();
789 if (NumOps > 1 && Desc.getOperandConstraint(1, TOI::TIED_TO) != -1)
791 else if (NumOps > 2 && Desc.getOperandConstraint(NumOps-1, TOI::TIED_TO)== 0)
792 // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32
795 // Keep track of the current byte being emitted.
796 unsigned CurByte = 0;
798 // Is this instruction encoded using the AVX VEX prefix?
799 bool HasVEXPrefix = false;
801 // It uses the VEX.VVVV field?
802 bool HasVEX_4V = false;
804 if ((TSFlags >> 32) & X86II::VEX)
806 if ((TSFlags >> 32) & X86II::VEX_4V)
809 // Determine where the memory operand starts, if present.
810 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags);
811 if (MemoryOperand != -1) MemoryOperand += CurOp;
814 EmitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
816 EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
818 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
819 unsigned SrcRegNum = 0;
820 switch (TSFlags & X86II::FormMask) {
821 case X86II::MRMInitReg:
822 assert(0 && "FIXME: Remove this form when the JIT moves to MCCodeEmitter!");
823 default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
824 assert(0 && "Unknown FormMask value in X86MCCodeEmitter!");
826 assert(0 && "Pseudo instruction shouldn't be emitted");
828 EmitByte(BaseOpcode, CurByte, OS);
831 case X86II::RawFrmImm16:
832 EmitByte(BaseOpcode, CurByte, OS);
833 EmitImmediate(MI.getOperand(CurOp++),
834 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
835 CurByte, OS, Fixups);
836 EmitImmediate(MI.getOperand(CurOp++), 2, FK_Data_2, CurByte, OS, Fixups);
839 case X86II::AddRegFrm:
840 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
843 case X86II::MRMDestReg:
844 EmitByte(BaseOpcode, CurByte, OS);
845 EmitRegModRMByte(MI.getOperand(CurOp),
846 GetX86RegNum(MI.getOperand(CurOp+1)), CurByte, OS);
850 case X86II::MRMDestMem:
851 EmitByte(BaseOpcode, CurByte, OS);
852 SrcRegNum = CurOp + X86::AddrNumOperands;
854 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
857 EmitMemModRMByte(MI, CurOp,
858 GetX86RegNum(MI.getOperand(SrcRegNum)),
859 TSFlags, CurByte, OS, Fixups);
860 CurOp = SrcRegNum + 1;
863 case X86II::MRMSrcReg:
864 EmitByte(BaseOpcode, CurByte, OS);
865 SrcRegNum = CurOp + 1;
867 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
870 EmitRegModRMByte(MI.getOperand(SrcRegNum),
871 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
872 CurOp = SrcRegNum + 1;
875 case X86II::MRMSrcMem: {
876 int AddrOperands = X86::AddrNumOperands;
877 unsigned FirstMemOp = CurOp+1;
880 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
883 EmitByte(BaseOpcode, CurByte, OS);
885 EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
886 TSFlags, CurByte, OS, Fixups);
887 CurOp += AddrOperands + 1;
891 case X86II::MRM0r: case X86II::MRM1r:
892 case X86II::MRM2r: case X86II::MRM3r:
893 case X86II::MRM4r: case X86II::MRM5r:
894 case X86II::MRM6r: case X86II::MRM7r:
895 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
897 EmitByte(BaseOpcode, CurByte, OS);
898 EmitRegModRMByte(MI.getOperand(CurOp++),
899 (TSFlags & X86II::FormMask)-X86II::MRM0r,
902 case X86II::MRM0m: case X86II::MRM1m:
903 case X86II::MRM2m: case X86II::MRM3m:
904 case X86II::MRM4m: case X86II::MRM5m:
905 case X86II::MRM6m: case X86II::MRM7m:
906 EmitByte(BaseOpcode, CurByte, OS);
907 EmitMemModRMByte(MI, CurOp, (TSFlags & X86II::FormMask)-X86II::MRM0m,
908 TSFlags, CurByte, OS, Fixups);
909 CurOp += X86::AddrNumOperands;
912 EmitByte(BaseOpcode, CurByte, OS);
913 EmitByte(0xC1, CurByte, OS);
916 EmitByte(BaseOpcode, CurByte, OS);
917 EmitByte(0xC2, CurByte, OS);
920 EmitByte(BaseOpcode, CurByte, OS);
921 EmitByte(0xC3, CurByte, OS);
924 EmitByte(BaseOpcode, CurByte, OS);
925 EmitByte(0xC4, CurByte, OS);
928 EmitByte(BaseOpcode, CurByte, OS);
929 EmitByte(0xC8, CurByte, OS);
932 EmitByte(BaseOpcode, CurByte, OS);
933 EmitByte(0xC9, CurByte, OS);
936 EmitByte(BaseOpcode, CurByte, OS);
937 EmitByte(0xE8, CurByte, OS);
940 EmitByte(BaseOpcode, CurByte, OS);
941 EmitByte(0xF0, CurByte, OS);
944 EmitByte(BaseOpcode, CurByte, OS);
945 EmitByte(0xF8, CurByte, OS);
948 EmitByte(BaseOpcode, CurByte, OS);
949 EmitByte(0xF9, CurByte, OS);
953 // If there is a remaining operand, it must be a trailing immediate. Emit it
954 // according to the right size for the instruction.
955 if (CurOp != NumOps) {
956 // The last source register of a 4 operand instruction in AVX is encoded
957 // in bits[7:4] of a immediate byte, and bits[3:0] are ignored.
958 if ((TSFlags >> 32) & X86II::VEX_I8IMM) {
959 const MCOperand &MO = MI.getOperand(CurOp++);
961 X86InstrInfo::isX86_64ExtendedReg(MO.getReg());
962 unsigned RegNum = (IsExtReg ? (1 << 7) : 0);
963 RegNum |= GetX86RegNum(MO) << 4;
964 EmitImmediate(MCOperand::CreateImm(RegNum), 1, FK_Data_1, CurByte, OS,
967 EmitImmediate(MI.getOperand(CurOp++),
968 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
969 CurByte, OS, Fixups);
975 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
976 errs() << "Cannot encode all operands of: ";