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;
182 /// Is32BitMemOperand - Return true if the specified instruction with a memory
183 /// operand should emit the 0x67 prefix byte in 64-bit mode due to a 32-bit
184 /// memory operand. Op specifies the operand # of the memoperand.
185 static bool Is32BitMemOperand(const MCInst &MI, unsigned Op) {
186 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
187 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
189 if ((BaseReg.getReg() != 0 && X86::GR32RegClass.contains(BaseReg.getReg())) ||
190 (IndexReg.getReg() != 0 && X86::GR32RegClass.contains(IndexReg.getReg())))
195 void X86MCCodeEmitter::
196 EmitImmediate(const MCOperand &DispOp, unsigned Size, MCFixupKind FixupKind,
197 unsigned &CurByte, raw_ostream &OS,
198 SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
199 // If this is a simple integer displacement that doesn't require a relocation,
201 if (DispOp.isImm()) {
202 // FIXME: is this right for pc-rel encoding?? Probably need to emit this as
204 EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
208 // If we have an immoffset, add it to the expression.
209 const MCExpr *Expr = DispOp.getExpr();
211 // If the fixup is pc-relative, we need to bias the value to be relative to
212 // the start of the field, not the end of the field.
213 if (FixupKind == MCFixupKind(X86::reloc_pcrel_4byte) ||
214 FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
215 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load))
217 if (FixupKind == MCFixupKind(X86::reloc_pcrel_2byte))
219 if (FixupKind == MCFixupKind(X86::reloc_pcrel_1byte))
223 Expr = MCBinaryExpr::CreateAdd(Expr, MCConstantExpr::Create(ImmOffset, Ctx),
226 // Emit a symbolic constant as a fixup and 4 zeros.
227 Fixups.push_back(MCFixup::Create(CurByte, Expr, FixupKind));
228 EmitConstant(0, Size, CurByte, OS);
231 void X86MCCodeEmitter::EmitMemModRMByte(const MCInst &MI, unsigned Op,
232 unsigned RegOpcodeField,
233 uint64_t TSFlags, unsigned &CurByte,
235 SmallVectorImpl<MCFixup> &Fixups) const{
236 const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
237 const MCOperand &Base = MI.getOperand(Op+X86::AddrBaseReg);
238 const MCOperand &Scale = MI.getOperand(Op+X86::AddrScaleAmt);
239 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
240 unsigned BaseReg = Base.getReg();
242 // Handle %rip relative addressing.
243 if (BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
244 assert(Is64BitMode && "Rip-relative addressing requires 64-bit mode");
245 assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
246 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
248 unsigned FixupKind = X86::reloc_riprel_4byte;
250 // movq loads are handled with a special relocation form which allows the
251 // linker to eliminate some loads for GOT references which end up in the
252 // same linkage unit.
253 if (MI.getOpcode() == X86::MOV64rm ||
254 MI.getOpcode() == X86::MOV64rm_TC)
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, FK_Data_4, CurByte, OS, Fixups);
314 // We need a SIB byte, so start by outputting the ModR/M byte first
315 assert(IndexReg.getReg() != X86::ESP &&
316 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
318 bool ForceDisp32 = false;
319 bool ForceDisp8 = false;
321 // If there is no base register, we emit the special case SIB byte with
322 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
323 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
325 } else if (!Disp.isImm()) {
326 // Emit the normal disp32 encoding.
327 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
329 } else if (Disp.getImm() == 0 &&
330 // Base reg can't be anything that ends up with '5' as the base
331 // reg, it is the magic [*] nomenclature that indicates no base.
332 BaseRegNo != N86::EBP) {
333 // Emit no displacement ModR/M byte
334 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
335 } else if (isDisp8(Disp.getImm())) {
336 // Emit the disp8 encoding.
337 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
338 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
340 // Emit the normal disp32 encoding.
341 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
344 // Calculate what the SS field value should be...
345 static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 };
346 unsigned SS = SSTable[Scale.getImm()];
349 // Handle the SIB byte for the case where there is no base, see Intel
350 // Manual 2A, table 2-7. The displacement has already been output.
352 if (IndexReg.getReg())
353 IndexRegNo = GetX86RegNum(IndexReg);
354 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
356 EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
359 if (IndexReg.getReg())
360 IndexRegNo = GetX86RegNum(IndexReg);
362 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
363 EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
366 // Do we need to output a displacement?
368 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
369 else if (ForceDisp32 || Disp.getImm() != 0)
370 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
373 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
375 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
376 int MemOperand, const MCInst &MI,
377 const TargetInstrDesc &Desc,
378 raw_ostream &OS) const {
379 bool HasVEX_4V = false;
380 if ((TSFlags >> 32) & X86II::VEX_4V)
383 // VEX_R: opcode externsion equivalent to REX.R in
384 // 1's complement (inverted) form
386 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
387 // 0: Same as REX_R=1 (64 bit mode only)
389 unsigned char VEX_R = 0x1;
391 // VEX_X: equivalent to REX.X, only used when a
392 // register is used for index in SIB Byte.
394 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
395 // 0: Same as REX.X=1 (64-bit mode only)
396 unsigned char VEX_X = 0x1;
400 // 1: Same as REX_B=0 (ignored in 32-bit mode)
401 // 0: Same as REX_B=1 (64 bit mode only)
403 unsigned char VEX_B = 0x1;
405 // VEX_W: opcode specific (use like REX.W, or used for
406 // opcode extension, or ignored, depending on the opcode byte)
407 unsigned char VEX_W = 0;
409 // VEX_5M (VEX m-mmmmm field):
411 // 0b00000: Reserved for future use
412 // 0b00001: implied 0F leading opcode
413 // 0b00010: implied 0F 38 leading opcode bytes
414 // 0b00011: implied 0F 3A leading opcode bytes
415 // 0b00100-0b11111: Reserved for future use
417 unsigned char VEX_5M = 0x1;
419 // VEX_4V (VEX vvvv field): a register specifier
420 // (in 1's complement form) or 1111 if unused.
421 unsigned char VEX_4V = 0xf;
423 // VEX_L (Vector Length):
425 // 0: scalar or 128-bit vector
428 unsigned char VEX_L = 0;
430 // VEX_PP: opcode extension providing equivalent
431 // functionality of a SIMD prefix
438 unsigned char VEX_PP = 0;
440 // Encode the operand size opcode prefix as needed.
441 if (TSFlags & X86II::OpSize)
444 if ((TSFlags >> 32) & X86II::VEX_W)
447 if ((TSFlags >> 32) & X86II::VEX_L)
450 switch (TSFlags & X86II::Op0Mask) {
451 default: assert(0 && "Invalid prefix!");
452 case X86II::T8: // 0F 38
455 case X86II::TA: // 0F 3A
458 case X86II::TF: // F2 0F 38
462 case X86II::XS: // F3 0F
465 case X86II::XD: // F2 0F
468 case X86II::TB: // Bypass: Not used by VEX
473 // Set the vector length to 256-bit if YMM0-YMM15 is used
474 for (unsigned i = 0; i != MI.getNumOperands(); ++i) {
475 if (!MI.getOperand(i).isReg())
477 unsigned SrcReg = MI.getOperand(i).getReg();
478 if (SrcReg >= X86::YMM0 && SrcReg <= X86::YMM15)
482 unsigned NumOps = MI.getNumOperands();
484 bool IsDestMem = false;
486 switch (TSFlags & X86II::FormMask) {
487 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
488 case X86II::MRMDestMem:
490 // The important info for the VEX prefix is never beyond the address
491 // registers. Don't check beyond that.
492 NumOps = CurOp = X86::AddrNumOperands;
493 case X86II::MRM0m: case X86II::MRM1m:
494 case X86II::MRM2m: case X86II::MRM3m:
495 case X86II::MRM4m: case X86II::MRM5m:
496 case X86II::MRM6m: case X86II::MRM7m:
497 case X86II::MRMSrcMem:
498 case X86II::MRMSrcReg:
499 if (MI.getNumOperands() > CurOp && MI.getOperand(CurOp).isReg() &&
500 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
505 VEX_4V = getVEXRegisterEncoding(MI, IsDestMem ? CurOp-1 : CurOp);
509 // To only check operands before the memory address ones, start
510 // the search from the begining
514 // If the last register should be encoded in the immediate field
515 // do not use any bit from VEX prefix to this register, ignore it
516 if ((TSFlags >> 32) & X86II::VEX_I8IMM)
519 for (; CurOp != NumOps; ++CurOp) {
520 const MCOperand &MO = MI.getOperand(CurOp);
521 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
523 if (!VEX_B && MO.isReg() &&
524 ((TSFlags & X86II::FormMask) == X86II::MRMSrcMem) &&
525 X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
529 default: // MRMDestReg, MRM0r-MRM7r, RawFrm
530 if (!MI.getNumOperands())
533 if (MI.getOperand(CurOp).isReg() &&
534 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
538 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
541 for (; CurOp != NumOps; ++CurOp) {
542 const MCOperand &MO = MI.getOperand(CurOp);
543 if (MO.isReg() && !HasVEX_4V &&
544 X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
550 // Emit segment override opcode prefix as needed.
551 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
553 // VEX opcode prefix can have 2 or 3 bytes
556 // +-----+ +--------------+ +-------------------+
557 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
558 // +-----+ +--------------+ +-------------------+
560 // +-----+ +-------------------+
561 // | C5h | | R | vvvv | L | pp |
562 // +-----+ +-------------------+
564 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
566 if (VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) { // 2 byte VEX prefix
567 EmitByte(0xC5, CurByte, OS);
568 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
573 EmitByte(0xC4, CurByte, OS);
574 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
575 EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
578 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
579 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
580 /// size, and 3) use of X86-64 extended registers.
581 static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
582 const TargetInstrDesc &Desc) {
584 if (TSFlags & X86II::REX_W)
585 REX |= 1 << 3; // set REX.W
587 if (MI.getNumOperands() == 0) return REX;
589 unsigned NumOps = MI.getNumOperands();
590 // FIXME: MCInst should explicitize the two-addrness.
591 bool isTwoAddr = NumOps > 1 &&
592 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
594 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
595 unsigned i = isTwoAddr ? 1 : 0;
596 for (; i != NumOps; ++i) {
597 const MCOperand &MO = MI.getOperand(i);
598 if (!MO.isReg()) continue;
599 unsigned Reg = MO.getReg();
600 if (!X86InstrInfo::isX86_64NonExtLowByteReg(Reg)) continue;
601 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
602 // that returns non-zero.
603 REX |= 0x40; // REX fixed encoding prefix
607 switch (TSFlags & X86II::FormMask) {
608 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
609 case X86II::MRMSrcReg:
610 if (MI.getOperand(0).isReg() &&
611 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
612 REX |= 1 << 2; // set REX.R
613 i = isTwoAddr ? 2 : 1;
614 for (; i != NumOps; ++i) {
615 const MCOperand &MO = MI.getOperand(i);
616 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
617 REX |= 1 << 0; // set REX.B
620 case X86II::MRMSrcMem: {
621 if (MI.getOperand(0).isReg() &&
622 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
623 REX |= 1 << 2; // set REX.R
625 i = isTwoAddr ? 2 : 1;
626 for (; i != NumOps; ++i) {
627 const MCOperand &MO = MI.getOperand(i);
629 if (X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
630 REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1)
636 case X86II::MRM0m: case X86II::MRM1m:
637 case X86II::MRM2m: case X86II::MRM3m:
638 case X86II::MRM4m: case X86II::MRM5m:
639 case X86II::MRM6m: case X86II::MRM7m:
640 case X86II::MRMDestMem: {
641 unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands);
642 i = isTwoAddr ? 1 : 0;
643 if (NumOps > e && MI.getOperand(e).isReg() &&
644 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(e).getReg()))
645 REX |= 1 << 2; // set REX.R
647 for (; i != e; ++i) {
648 const MCOperand &MO = MI.getOperand(i);
650 if (X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
651 REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1)
658 if (MI.getOperand(0).isReg() &&
659 X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
660 REX |= 1 << 0; // set REX.B
661 i = isTwoAddr ? 2 : 1;
662 for (unsigned e = NumOps; i != e; ++i) {
663 const MCOperand &MO = MI.getOperand(i);
664 if (MO.isReg() && X86InstrInfo::isX86_64ExtendedReg(MO.getReg()))
665 REX |= 1 << 2; // set REX.R
672 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
673 void X86MCCodeEmitter::EmitSegmentOverridePrefix(uint64_t TSFlags,
674 unsigned &CurByte, int MemOperand,
676 raw_ostream &OS) const {
677 switch (TSFlags & X86II::SegOvrMask) {
678 default: assert(0 && "Invalid segment!");
680 // No segment override, check for explicit one on memory operand.
681 if (MemOperand != -1) { // If the instruction has a memory operand.
682 switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
683 default: assert(0 && "Unknown segment register!");
685 case X86::CS: EmitByte(0x2E, CurByte, OS); break;
686 case X86::SS: EmitByte(0x36, CurByte, OS); break;
687 case X86::DS: EmitByte(0x3E, CurByte, OS); break;
688 case X86::ES: EmitByte(0x26, CurByte, OS); break;
689 case X86::FS: EmitByte(0x64, CurByte, OS); break;
690 case X86::GS: EmitByte(0x65, CurByte, OS); break;
695 EmitByte(0x64, CurByte, OS);
698 EmitByte(0x65, CurByte, OS);
703 /// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode.
705 /// MemOperand is the operand # of the start of a memory operand if present. If
706 /// Not present, it is -1.
707 void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
708 int MemOperand, const MCInst &MI,
709 const TargetInstrDesc &Desc,
710 raw_ostream &OS) const {
712 // Emit the lock opcode prefix as needed.
713 if (TSFlags & X86II::LOCK)
714 EmitByte(0xF0, CurByte, OS);
716 // Emit segment override opcode prefix as needed.
717 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
719 // Emit the repeat opcode prefix as needed.
720 if ((TSFlags & X86II::Op0Mask) == X86II::REP)
721 EmitByte(0xF3, CurByte, OS);
723 // Emit the address size opcode prefix as needed.
724 if ((TSFlags & X86II::AdSize) ||
725 (MemOperand != -1 && Is64BitMode && Is32BitMemOperand(MI, MemOperand)))
726 EmitByte(0x67, CurByte, OS);
728 // Emit the operand size opcode prefix as needed.
729 if (TSFlags & X86II::OpSize)
730 EmitByte(0x66, CurByte, OS);
732 bool Need0FPrefix = false;
733 switch (TSFlags & X86II::Op0Mask) {
734 default: assert(0 && "Invalid prefix!");
735 case 0: break; // No prefix!
736 case X86II::REP: break; // already handled.
737 case X86II::TB: // Two-byte opcode prefix
738 case X86II::T8: // 0F 38
739 case X86II::TA: // 0F 3A
742 case X86II::TF: // F2 0F 38
743 EmitByte(0xF2, CurByte, OS);
746 case X86II::XS: // F3 0F
747 EmitByte(0xF3, CurByte, OS);
750 case X86II::XD: // F2 0F
751 EmitByte(0xF2, CurByte, OS);
754 case X86II::D8: EmitByte(0xD8, CurByte, OS); break;
755 case X86II::D9: EmitByte(0xD9, CurByte, OS); break;
756 case X86II::DA: EmitByte(0xDA, CurByte, OS); break;
757 case X86II::DB: EmitByte(0xDB, CurByte, OS); break;
758 case X86II::DC: EmitByte(0xDC, CurByte, OS); break;
759 case X86II::DD: EmitByte(0xDD, CurByte, OS); break;
760 case X86II::DE: EmitByte(0xDE, CurByte, OS); break;
761 case X86II::DF: EmitByte(0xDF, CurByte, OS); break;
764 // Handle REX prefix.
765 // FIXME: Can this come before F2 etc to simplify emission?
767 if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc))
768 EmitByte(0x40 | REX, CurByte, OS);
771 // 0x0F escape code must be emitted just before the opcode.
773 EmitByte(0x0F, CurByte, OS);
775 // FIXME: Pull this up into previous switch if REX can be moved earlier.
776 switch (TSFlags & X86II::Op0Mask) {
777 case X86II::TF: // F2 0F 38
778 case X86II::T8: // 0F 38
779 EmitByte(0x38, CurByte, OS);
781 case X86II::TA: // 0F 3A
782 EmitByte(0x3A, CurByte, OS);
787 void X86MCCodeEmitter::
788 EncodeInstruction(const MCInst &MI, raw_ostream &OS,
789 SmallVectorImpl<MCFixup> &Fixups) const {
790 unsigned Opcode = MI.getOpcode();
791 const TargetInstrDesc &Desc = TII.get(Opcode);
792 uint64_t TSFlags = Desc.TSFlags;
794 // Pseudo instructions don't get encoded.
795 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
798 // If this is a two-address instruction, skip one of the register operands.
799 // FIXME: This should be handled during MCInst lowering.
800 unsigned NumOps = Desc.getNumOperands();
802 if (NumOps > 1 && Desc.getOperandConstraint(1, TOI::TIED_TO) != -1)
804 else if (NumOps > 2 && Desc.getOperandConstraint(NumOps-1, TOI::TIED_TO)== 0)
805 // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32
808 // Keep track of the current byte being emitted.
809 unsigned CurByte = 0;
811 // Is this instruction encoded using the AVX VEX prefix?
812 bool HasVEXPrefix = false;
814 // It uses the VEX.VVVV field?
815 bool HasVEX_4V = false;
817 if ((TSFlags >> 32) & X86II::VEX)
819 if ((TSFlags >> 32) & X86II::VEX_4V)
822 // Determine where the memory operand starts, if present.
823 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags);
824 if (MemoryOperand != -1) MemoryOperand += CurOp;
827 EmitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
829 EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
831 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
832 unsigned SrcRegNum = 0;
833 switch (TSFlags & X86II::FormMask) {
834 case X86II::MRMInitReg:
835 assert(0 && "FIXME: Remove this form when the JIT moves to MCCodeEmitter!");
836 default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
837 assert(0 && "Unknown FormMask value in X86MCCodeEmitter!");
839 assert(0 && "Pseudo instruction shouldn't be emitted");
841 EmitByte(BaseOpcode, CurByte, OS);
844 case X86II::RawFrmImm8:
845 EmitByte(BaseOpcode, CurByte, OS);
846 EmitImmediate(MI.getOperand(CurOp++),
847 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
848 CurByte, OS, Fixups);
849 EmitImmediate(MI.getOperand(CurOp++), 1, FK_Data_1, CurByte, OS, Fixups);
851 case X86II::RawFrmImm16:
852 EmitByte(BaseOpcode, CurByte, OS);
853 EmitImmediate(MI.getOperand(CurOp++),
854 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
855 CurByte, OS, Fixups);
856 EmitImmediate(MI.getOperand(CurOp++), 2, FK_Data_2, CurByte, OS, Fixups);
859 case X86II::AddRegFrm:
860 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
863 case X86II::MRMDestReg:
864 EmitByte(BaseOpcode, CurByte, OS);
865 EmitRegModRMByte(MI.getOperand(CurOp),
866 GetX86RegNum(MI.getOperand(CurOp+1)), CurByte, OS);
870 case X86II::MRMDestMem:
871 EmitByte(BaseOpcode, CurByte, OS);
872 SrcRegNum = CurOp + X86::AddrNumOperands;
874 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
877 EmitMemModRMByte(MI, CurOp,
878 GetX86RegNum(MI.getOperand(SrcRegNum)),
879 TSFlags, CurByte, OS, Fixups);
880 CurOp = SrcRegNum + 1;
883 case X86II::MRMSrcReg:
884 EmitByte(BaseOpcode, CurByte, OS);
885 SrcRegNum = CurOp + 1;
887 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
890 EmitRegModRMByte(MI.getOperand(SrcRegNum),
891 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
892 CurOp = SrcRegNum + 1;
895 case X86II::MRMSrcMem: {
896 int AddrOperands = X86::AddrNumOperands;
897 unsigned FirstMemOp = CurOp+1;
900 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
903 EmitByte(BaseOpcode, CurByte, OS);
905 EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
906 TSFlags, CurByte, OS, Fixups);
907 CurOp += AddrOperands + 1;
911 case X86II::MRM0r: case X86II::MRM1r:
912 case X86II::MRM2r: case X86II::MRM3r:
913 case X86II::MRM4r: case X86II::MRM5r:
914 case X86II::MRM6r: case X86II::MRM7r:
915 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
917 EmitByte(BaseOpcode, CurByte, OS);
918 EmitRegModRMByte(MI.getOperand(CurOp++),
919 (TSFlags & X86II::FormMask)-X86II::MRM0r,
922 case X86II::MRM0m: case X86II::MRM1m:
923 case X86II::MRM2m: case X86II::MRM3m:
924 case X86II::MRM4m: case X86II::MRM5m:
925 case X86II::MRM6m: case X86II::MRM7m:
926 EmitByte(BaseOpcode, CurByte, OS);
927 EmitMemModRMByte(MI, CurOp, (TSFlags & X86II::FormMask)-X86II::MRM0m,
928 TSFlags, CurByte, OS, Fixups);
929 CurOp += X86::AddrNumOperands;
932 EmitByte(BaseOpcode, CurByte, OS);
933 EmitByte(0xC1, CurByte, OS);
936 EmitByte(BaseOpcode, CurByte, OS);
937 EmitByte(0xC2, CurByte, OS);
940 EmitByte(BaseOpcode, CurByte, OS);
941 EmitByte(0xC3, CurByte, OS);
944 EmitByte(BaseOpcode, CurByte, OS);
945 EmitByte(0xC4, CurByte, OS);
948 EmitByte(BaseOpcode, CurByte, OS);
949 EmitByte(0xC8, CurByte, OS);
952 EmitByte(BaseOpcode, CurByte, OS);
953 EmitByte(0xC9, CurByte, OS);
956 EmitByte(BaseOpcode, CurByte, OS);
957 EmitByte(0xE8, CurByte, OS);
960 EmitByte(BaseOpcode, CurByte, OS);
961 EmitByte(0xF0, CurByte, OS);
964 EmitByte(BaseOpcode, CurByte, OS);
965 EmitByte(0xF8, CurByte, OS);
968 EmitByte(BaseOpcode, CurByte, OS);
969 EmitByte(0xF9, CurByte, OS);
973 // If there is a remaining operand, it must be a trailing immediate. Emit it
974 // according to the right size for the instruction.
975 if (CurOp != NumOps) {
976 // The last source register of a 4 operand instruction in AVX is encoded
977 // in bits[7:4] of a immediate byte, and bits[3:0] are ignored.
978 if ((TSFlags >> 32) & X86II::VEX_I8IMM) {
979 const MCOperand &MO = MI.getOperand(CurOp++);
981 X86InstrInfo::isX86_64ExtendedReg(MO.getReg());
982 unsigned RegNum = (IsExtReg ? (1 << 7) : 0);
983 RegNum |= GetX86RegNum(MO) << 4;
984 EmitImmediate(MCOperand::CreateImm(RegNum), 1, FK_Data_1, CurByte, OS,
987 EmitImmediate(MI.getOperand(CurOp++),
988 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
989 CurByte, OS, Fixups);
995 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
996 errs() << "Cannot encode all operands of: ";