1 //===-- 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 #include "MCTargetDesc/X86MCTargetDesc.h"
15 #include "MCTargetDesc/X86BaseInfo.h"
16 #include "MCTargetDesc/X86FixupKinds.h"
17 #include "llvm/MC/MCCodeEmitter.h"
18 #include "llvm/MC/MCContext.h"
19 #include "llvm/MC/MCExpr.h"
20 #include "llvm/MC/MCInst.h"
21 #include "llvm/MC/MCInstrInfo.h"
22 #include "llvm/MC/MCRegisterInfo.h"
23 #include "llvm/MC/MCSubtargetInfo.h"
24 #include "llvm/MC/MCSymbol.h"
25 #include "llvm/Support/raw_ostream.h"
29 #define DEBUG_TYPE "mccodeemitter"
32 class X86MCCodeEmitter : public MCCodeEmitter {
33 X86MCCodeEmitter(const X86MCCodeEmitter &) LLVM_DELETED_FUNCTION;
34 void operator=(const X86MCCodeEmitter &) LLVM_DELETED_FUNCTION;
35 const MCInstrInfo &MCII;
38 X86MCCodeEmitter(const MCInstrInfo &mcii, MCContext &ctx)
39 : MCII(mcii), Ctx(ctx) {
42 ~X86MCCodeEmitter() {}
44 bool is64BitMode(const MCSubtargetInfo &STI) const {
45 return (STI.getFeatureBits() & X86::Mode64Bit) != 0;
48 bool is32BitMode(const MCSubtargetInfo &STI) const {
49 return (STI.getFeatureBits() & X86::Mode32Bit) != 0;
52 bool is16BitMode(const MCSubtargetInfo &STI) const {
53 return (STI.getFeatureBits() & X86::Mode16Bit) != 0;
56 /// Is16BitMemOperand - Return true if the specified instruction has
57 /// a 16-bit memory operand. Op specifies the operand # of the memoperand.
58 bool Is16BitMemOperand(const MCInst &MI, unsigned Op,
59 const MCSubtargetInfo &STI) const {
60 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
61 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
62 const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
64 if (is16BitMode(STI) && BaseReg.getReg() == 0 &&
65 Disp.isImm() && Disp.getImm() < 0x10000)
67 if ((BaseReg.getReg() != 0 &&
68 X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) ||
69 (IndexReg.getReg() != 0 &&
70 X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg())))
75 unsigned GetX86RegNum(const MCOperand &MO) const {
76 return Ctx.getRegisterInfo()->getEncodingValue(MO.getReg()) & 0x7;
79 // On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the range
80 // 0-7 and the difference between the 2 groups is given by the REX prefix.
81 // In the VEX prefix, registers are seen sequencially from 0-15 and encoded
82 // in 1's complement form, example:
84 // ModRM field => XMM9 => 1
85 // VEX.VVVV => XMM9 => ~9
87 // See table 4-35 of Intel AVX Programming Reference for details.
88 unsigned char getVEXRegisterEncoding(const MCInst &MI,
89 unsigned OpNum) const {
90 unsigned SrcReg = MI.getOperand(OpNum).getReg();
91 unsigned SrcRegNum = GetX86RegNum(MI.getOperand(OpNum));
92 if (X86II::isX86_64ExtendedReg(SrcReg))
95 // The registers represented through VEX_VVVV should
96 // be encoded in 1's complement form.
97 return (~SrcRegNum) & 0xf;
100 unsigned char getWriteMaskRegisterEncoding(const MCInst &MI,
101 unsigned OpNum) const {
102 assert(X86::K0 != MI.getOperand(OpNum).getReg() &&
103 "Invalid mask register as write-mask!");
104 unsigned MaskRegNum = GetX86RegNum(MI.getOperand(OpNum));
108 void EmitByte(unsigned char C, unsigned &CurByte, raw_ostream &OS) const {
113 void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
114 raw_ostream &OS) const {
115 // Output the constant in little endian byte order.
116 for (unsigned i = 0; i != Size; ++i) {
117 EmitByte(Val & 255, CurByte, OS);
122 void EmitImmediate(const MCOperand &Disp, SMLoc Loc,
123 unsigned ImmSize, MCFixupKind FixupKind,
124 unsigned &CurByte, raw_ostream &OS,
125 SmallVectorImpl<MCFixup> &Fixups,
126 int ImmOffset = 0) const;
128 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
130 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
131 return RM | (RegOpcode << 3) | (Mod << 6);
134 void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
135 unsigned &CurByte, raw_ostream &OS) const {
136 EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS);
139 void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base,
140 unsigned &CurByte, raw_ostream &OS) const {
141 // SIB byte is in the same format as the ModRMByte.
142 EmitByte(ModRMByte(SS, Index, Base), CurByte, OS);
146 void EmitMemModRMByte(const MCInst &MI, unsigned Op,
147 unsigned RegOpcodeField,
148 uint64_t TSFlags, unsigned &CurByte, raw_ostream &OS,
149 SmallVectorImpl<MCFixup> &Fixups,
150 const MCSubtargetInfo &STI) const;
152 void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
153 SmallVectorImpl<MCFixup> &Fixups,
154 const MCSubtargetInfo &STI) const override;
156 void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
157 const MCInst &MI, const MCInstrDesc &Desc,
158 raw_ostream &OS) const;
160 void EmitSegmentOverridePrefix(unsigned &CurByte, unsigned SegOperand,
161 const MCInst &MI, raw_ostream &OS) const;
163 void EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
164 const MCInst &MI, const MCInstrDesc &Desc,
165 const MCSubtargetInfo &STI,
166 raw_ostream &OS) const;
169 } // end anonymous namespace
172 MCCodeEmitter *llvm::createX86MCCodeEmitter(const MCInstrInfo &MCII,
173 const MCRegisterInfo &MRI,
174 const MCSubtargetInfo &STI,
176 return new X86MCCodeEmitter(MCII, Ctx);
179 /// isDisp8 - Return true if this signed displacement fits in a 8-bit
180 /// sign-extended field.
181 static bool isDisp8(int Value) {
182 return Value == (signed char)Value;
185 /// isCDisp8 - Return true if this signed displacement fits in a 8-bit
186 /// compressed dispacement field.
187 static bool isCDisp8(uint64_t TSFlags, int Value, int& CValue) {
188 assert(((TSFlags & X86II::EncodingMask) >>
189 X86II::EncodingShift == X86II::EVEX) &&
190 "Compressed 8-bit displacement is only valid for EVEX inst.");
192 unsigned CD8E = (TSFlags >> X86II::EVEX_CD8EShift) & X86II::EVEX_CD8EMask;
193 unsigned CD8V = (TSFlags >> X86II::EVEX_CD8VShift) & X86II::EVEX_CD8VMask;
194 unsigned CD8_Scale = (TSFlags >> 56) & 0x7f;
196 if (CD8V == 0 && CD8E == 0) {
198 assert(CD8_Scale == 0);
199 return isDisp8(Value);
202 unsigned ElemSize = 1U << CD8E;
204 // The unit of displacement is either
205 // - the size of a power-of-two number of elements or
206 // - the size of a single element for broadcasts or
207 // - the total vector size divided by a power-of-two number.
209 // Fixed vector length
210 unsigned NumElems = 1U << (CD8V & 0x3);
211 MemObjSize = ElemSize * NumElems;
213 // Modified vector length
214 bool EVEX_b = (TSFlags >> X86II::VEXShift) & X86II::EVEX_B;
216 // Broadcast implies element size units.
217 MemObjSize = ElemSize;
219 unsigned EVEX_LL = ((TSFlags >> X86II::VEXShift) & X86II::VEX_L) ? 1 : 0;
220 EVEX_LL += ((TSFlags >> X86II::VEXShift) & X86II::EVEX_L2) ? 2 : 0;
221 assert(EVEX_LL < 3 && "");
223 unsigned VectorByteSize = 1U << (EVEX_LL + 4);
224 unsigned Divider = 1U << (CD8V & 0x3);
225 MemObjSize = VectorByteSize / Divider;
229 assert(MemObjSize == CD8_Scale);
231 unsigned MemObjMask = MemObjSize - 1;
232 assert((MemObjSize & MemObjMask) == 0 && "Invalid memory object size.");
234 if (Value & MemObjMask) // Unaligned offset
236 Value /= (int)MemObjSize;
237 bool Ret = (Value == (signed char)Value);
244 /// getImmFixupKind - Return the appropriate fixup kind to use for an immediate
245 /// in an instruction with the specified TSFlags.
246 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
247 unsigned Size = X86II::getSizeOfImm(TSFlags);
248 bool isPCRel = X86II::isImmPCRel(TSFlags);
250 if (X86II::isImmSigned(TSFlags)) {
252 default: llvm_unreachable("Unsupported signed fixup size!");
253 case 4: return MCFixupKind(X86::reloc_signed_4byte);
256 return MCFixup::getKindForSize(Size, isPCRel);
259 /// Is32BitMemOperand - Return true if the specified instruction has
260 /// a 32-bit memory operand. Op specifies the operand # of the memoperand.
261 static bool Is32BitMemOperand(const MCInst &MI, unsigned Op) {
262 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
263 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
265 if ((BaseReg.getReg() != 0 &&
266 X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) ||
267 (IndexReg.getReg() != 0 &&
268 X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg())))
273 /// Is64BitMemOperand - Return true if the specified instruction has
274 /// a 64-bit memory operand. Op specifies the operand # of the memoperand.
276 static bool Is64BitMemOperand(const MCInst &MI, unsigned Op) {
277 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
278 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
280 if ((BaseReg.getReg() != 0 &&
281 X86MCRegisterClasses[X86::GR64RegClassID].contains(BaseReg.getReg())) ||
282 (IndexReg.getReg() != 0 &&
283 X86MCRegisterClasses[X86::GR64RegClassID].contains(IndexReg.getReg())))
289 /// StartsWithGlobalOffsetTable - Check if this expression starts with
290 /// _GLOBAL_OFFSET_TABLE_ and if it is of the form
291 /// _GLOBAL_OFFSET_TABLE_-symbol. This is needed to support PIC on ELF
292 /// i386 as _GLOBAL_OFFSET_TABLE_ is magical. We check only simple case that
293 /// are know to be used: _GLOBAL_OFFSET_TABLE_ by itself or at the start
294 /// of a binary expression.
295 enum GlobalOffsetTableExprKind {
300 static GlobalOffsetTableExprKind
301 StartsWithGlobalOffsetTable(const MCExpr *Expr) {
302 const MCExpr *RHS = nullptr;
303 if (Expr->getKind() == MCExpr::Binary) {
304 const MCBinaryExpr *BE = static_cast<const MCBinaryExpr *>(Expr);
309 if (Expr->getKind() != MCExpr::SymbolRef)
312 const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
313 const MCSymbol &S = Ref->getSymbol();
314 if (S.getName() != "_GLOBAL_OFFSET_TABLE_")
316 if (RHS && RHS->getKind() == MCExpr::SymbolRef)
321 static bool HasSecRelSymbolRef(const MCExpr *Expr) {
322 if (Expr->getKind() == MCExpr::SymbolRef) {
323 const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
324 return Ref->getKind() == MCSymbolRefExpr::VK_SECREL;
329 void X86MCCodeEmitter::
330 EmitImmediate(const MCOperand &DispOp, SMLoc Loc, unsigned Size,
331 MCFixupKind FixupKind, unsigned &CurByte, raw_ostream &OS,
332 SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
333 const MCExpr *Expr = nullptr;
334 if (DispOp.isImm()) {
335 // If this is a simple integer displacement that doesn't require a
336 // relocation, emit it now.
337 if (FixupKind != FK_PCRel_1 &&
338 FixupKind != FK_PCRel_2 &&
339 FixupKind != FK_PCRel_4) {
340 EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
343 Expr = MCConstantExpr::Create(DispOp.getImm(), Ctx);
345 Expr = DispOp.getExpr();
348 // If we have an immoffset, add it to the expression.
349 if ((FixupKind == FK_Data_4 ||
350 FixupKind == FK_Data_8 ||
351 FixupKind == MCFixupKind(X86::reloc_signed_4byte))) {
352 GlobalOffsetTableExprKind Kind = StartsWithGlobalOffsetTable(Expr);
353 if (Kind != GOT_None) {
354 assert(ImmOffset == 0);
357 FixupKind = MCFixupKind(X86::reloc_global_offset_table8);
360 FixupKind = MCFixupKind(X86::reloc_global_offset_table);
363 if (Kind == GOT_Normal)
365 } else if (Expr->getKind() == MCExpr::SymbolRef) {
366 if (HasSecRelSymbolRef(Expr)) {
367 FixupKind = MCFixupKind(FK_SecRel_4);
369 } else if (Expr->getKind() == MCExpr::Binary) {
370 const MCBinaryExpr *Bin = static_cast<const MCBinaryExpr*>(Expr);
371 if (HasSecRelSymbolRef(Bin->getLHS())
372 || HasSecRelSymbolRef(Bin->getRHS())) {
373 FixupKind = MCFixupKind(FK_SecRel_4);
378 // If the fixup is pc-relative, we need to bias the value to be relative to
379 // the start of the field, not the end of the field.
380 if (FixupKind == FK_PCRel_4 ||
381 FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
382 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load))
384 if (FixupKind == FK_PCRel_2)
386 if (FixupKind == FK_PCRel_1)
390 Expr = MCBinaryExpr::CreateAdd(Expr, MCConstantExpr::Create(ImmOffset, Ctx),
393 // Emit a symbolic constant as a fixup and 4 zeros.
394 Fixups.push_back(MCFixup::Create(CurByte, Expr, FixupKind, Loc));
395 EmitConstant(0, Size, CurByte, OS);
398 void X86MCCodeEmitter::EmitMemModRMByte(const MCInst &MI, unsigned Op,
399 unsigned RegOpcodeField,
400 uint64_t TSFlags, unsigned &CurByte,
402 SmallVectorImpl<MCFixup> &Fixups,
403 const MCSubtargetInfo &STI) const{
404 const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
405 const MCOperand &Base = MI.getOperand(Op+X86::AddrBaseReg);
406 const MCOperand &Scale = MI.getOperand(Op+X86::AddrScaleAmt);
407 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
408 unsigned BaseReg = Base.getReg();
409 unsigned char Encoding = (TSFlags & X86II::EncodingMask) >>
410 X86II::EncodingShift;
411 bool HasEVEX = (Encoding == X86II::EVEX);
413 // Handle %rip relative addressing.
414 if (BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
415 assert(is64BitMode(STI) && "Rip-relative addressing requires 64-bit mode");
416 assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
417 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
419 unsigned FixupKind = X86::reloc_riprel_4byte;
421 // movq loads are handled with a special relocation form which allows the
422 // linker to eliminate some loads for GOT references which end up in the
423 // same linkage unit.
424 if (MI.getOpcode() == X86::MOV64rm)
425 FixupKind = X86::reloc_riprel_4byte_movq_load;
427 // rip-relative addressing is actually relative to the *next* instruction.
428 // Since an immediate can follow the mod/rm byte for an instruction, this
429 // means that we need to bias the immediate field of the instruction with
430 // the size of the immediate field. If we have this case, add it into the
431 // expression to emit.
432 int ImmSize = X86II::hasImm(TSFlags) ? X86II::getSizeOfImm(TSFlags) : 0;
434 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind),
435 CurByte, OS, Fixups, -ImmSize);
439 unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
441 // 16-bit addressing forms of the ModR/M byte have a different encoding for
442 // the R/M field and are far more limited in which registers can be used.
443 if (Is16BitMemOperand(MI, Op, STI)) {
445 // For 32-bit addressing, the row and column values in Table 2-2 are
446 // basically the same. It's AX/CX/DX/BX/SP/BP/SI/DI in that order, with
447 // some special cases. And GetX86RegNum reflects that numbering.
448 // For 16-bit addressing it's more fun, as shown in the SDM Vol 2A,
449 // Table 2-1 "16-Bit Addressing Forms with the ModR/M byte". We can only
450 // use SI/DI/BP/BX, which have "row" values 4-7 in no particular order,
451 // while values 0-3 indicate the allowed combinations (base+index) of
452 // those: 0 for BX+SI, 1 for BX+DI, 2 for BP+SI, 3 for BP+DI.
454 // R16Table[] is a lookup from the normal RegNo, to the row values from
455 // Table 2-1 for 16-bit addressing modes. Where zero means disallowed.
456 static const unsigned R16Table[] = { 0, 0, 0, 7, 0, 6, 4, 5 };
457 unsigned RMfield = R16Table[BaseRegNo];
459 assert(RMfield && "invalid 16-bit base register");
461 if (IndexReg.getReg()) {
462 unsigned IndexReg16 = R16Table[GetX86RegNum(IndexReg)];
464 assert(IndexReg16 && "invalid 16-bit index register");
465 // We must have one of SI/DI (4,5), and one of BP/BX (6,7).
466 assert(((IndexReg16 ^ RMfield) & 2) &&
467 "invalid 16-bit base/index register combination");
468 assert(Scale.getImm() == 1 &&
469 "invalid scale for 16-bit memory reference");
471 // Allow base/index to appear in either order (although GAS doesn't).
473 RMfield = (RMfield & 1) | ((7 - IndexReg16) << 1);
475 RMfield = (IndexReg16 & 1) | ((7 - RMfield) << 1);
478 if (Disp.isImm() && isDisp8(Disp.getImm())) {
479 if (Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
480 // There is no displacement; just the register.
481 EmitByte(ModRMByte(0, RegOpcodeField, RMfield), CurByte, OS);
484 // Use the [REG]+disp8 form, including for [BP] which cannot be encoded.
485 EmitByte(ModRMByte(1, RegOpcodeField, RMfield), CurByte, OS);
486 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
489 // This is the [REG]+disp16 case.
490 EmitByte(ModRMByte(2, RegOpcodeField, RMfield), CurByte, OS);
492 // There is no BaseReg; this is the plain [disp16] case.
493 EmitByte(ModRMByte(0, RegOpcodeField, 6), CurByte, OS);
496 // Emit 16-bit displacement for plain disp16 or [REG]+disp16 cases.
497 EmitImmediate(Disp, MI.getLoc(), 2, FK_Data_2, CurByte, OS, Fixups);
501 // Determine whether a SIB byte is needed.
502 // If no BaseReg, issue a RIP relative instruction only if the MCE can
503 // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
504 // 2-7) and absolute references.
506 if (// The SIB byte must be used if there is an index register.
507 IndexReg.getReg() == 0 &&
508 // The SIB byte must be used if the base is ESP/RSP/R12, all of which
509 // encode to an R/M value of 4, which indicates that a SIB byte is
511 BaseRegNo != N86::ESP &&
512 // If there is no base register and we're in 64-bit mode, we need a SIB
513 // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
514 (!is64BitMode(STI) || BaseReg != 0)) {
516 if (BaseReg == 0) { // [disp32] in X86-32 mode
517 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
518 EmitImmediate(Disp, MI.getLoc(), 4, FK_Data_4, CurByte, OS, Fixups);
522 // If the base is not EBP/ESP and there is no displacement, use simple
523 // indirect register encoding, this handles addresses like [EAX]. The
524 // encoding for [EBP] with no displacement means [disp32] so we handle it
525 // by emitting a displacement of 0 below.
526 if (Disp.isImm() && Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
527 EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
531 // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
533 if (!HasEVEX && isDisp8(Disp.getImm())) {
534 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
535 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
538 // Try EVEX compressed 8-bit displacement first; if failed, fall back to
539 // 32-bit displacement.
541 if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
542 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
543 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups,
544 CDisp8 - Disp.getImm());
549 // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
550 EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
551 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(X86::reloc_signed_4byte), CurByte, OS,
556 // We need a SIB byte, so start by outputting the ModR/M byte first
557 assert(IndexReg.getReg() != X86::ESP &&
558 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
560 bool ForceDisp32 = false;
561 bool ForceDisp8 = false;
565 // If there is no base register, we emit the special case SIB byte with
566 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
567 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
569 } else if (!Disp.isImm()) {
570 // Emit the normal disp32 encoding.
571 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
573 } else if (Disp.getImm() == 0 &&
574 // Base reg can't be anything that ends up with '5' as the base
575 // reg, it is the magic [*] nomenclature that indicates no base.
576 BaseRegNo != N86::EBP) {
577 // Emit no displacement ModR/M byte
578 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
579 } else if (!HasEVEX && isDisp8(Disp.getImm())) {
580 // Emit the disp8 encoding.
581 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
582 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
583 } else if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
584 // Emit the disp8 encoding.
585 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
586 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
587 ImmOffset = CDisp8 - Disp.getImm();
589 // Emit the normal disp32 encoding.
590 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
593 // Calculate what the SS field value should be...
594 static const unsigned SSTable[] = { ~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3 };
595 unsigned SS = SSTable[Scale.getImm()];
598 // Handle the SIB byte for the case where there is no base, see Intel
599 // Manual 2A, table 2-7. The displacement has already been output.
601 if (IndexReg.getReg())
602 IndexRegNo = GetX86RegNum(IndexReg);
603 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
605 EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
608 if (IndexReg.getReg())
609 IndexRegNo = GetX86RegNum(IndexReg);
611 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
612 EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
615 // Do we need to output a displacement?
617 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups, ImmOffset);
618 else if (ForceDisp32 || Disp.getImm() != 0)
619 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(X86::reloc_signed_4byte),
620 CurByte, OS, Fixups);
623 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
625 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
626 int MemOperand, const MCInst &MI,
627 const MCInstrDesc &Desc,
628 raw_ostream &OS) const {
629 unsigned char Encoding = (TSFlags & X86II::EncodingMask) >>
630 X86II::EncodingShift;
631 bool HasEVEX_K = ((TSFlags >> X86II::VEXShift) & X86II::EVEX_K);
632 bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
633 bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
634 bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4;
635 bool HasEVEX_RC = (TSFlags >> X86II::VEXShift) & X86II::EVEX_RC;
637 // VEX_R: opcode externsion equivalent to REX.R in
638 // 1's complement (inverted) form
640 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
641 // 0: Same as REX_R=1 (64 bit mode only)
643 unsigned char VEX_R = 0x1;
644 unsigned char EVEX_R2 = 0x1;
646 // VEX_X: equivalent to REX.X, only used when a
647 // register is used for index in SIB Byte.
649 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
650 // 0: Same as REX.X=1 (64-bit mode only)
651 unsigned char VEX_X = 0x1;
655 // 1: Same as REX_B=0 (ignored in 32-bit mode)
656 // 0: Same as REX_B=1 (64 bit mode only)
658 unsigned char VEX_B = 0x1;
660 // VEX_W: opcode specific (use like REX.W, or used for
661 // opcode extension, or ignored, depending on the opcode byte)
662 unsigned char VEX_W = 0;
664 // VEX_5M (VEX m-mmmmm field):
666 // 0b00000: Reserved for future use
667 // 0b00001: implied 0F leading opcode
668 // 0b00010: implied 0F 38 leading opcode bytes
669 // 0b00011: implied 0F 3A leading opcode bytes
670 // 0b00100-0b11111: Reserved for future use
671 // 0b01000: XOP map select - 08h instructions with imm byte
672 // 0b01001: XOP map select - 09h instructions with no imm byte
673 // 0b01010: XOP map select - 0Ah instructions with imm dword
674 unsigned char VEX_5M = 0;
676 // VEX_4V (VEX vvvv field): a register specifier
677 // (in 1's complement form) or 1111 if unused.
678 unsigned char VEX_4V = 0xf;
679 unsigned char EVEX_V2 = 0x1;
681 // VEX_L (Vector Length):
683 // 0: scalar or 128-bit vector
686 unsigned char VEX_L = 0;
687 unsigned char EVEX_L2 = 0;
689 // VEX_PP: opcode extension providing equivalent
690 // functionality of a SIMD prefix
697 unsigned char VEX_PP = 0;
700 unsigned char EVEX_U = 1; // Always '1' so far
703 unsigned char EVEX_z = 0;
706 unsigned char EVEX_b = 0;
709 unsigned char EVEX_rc = 0;
712 unsigned char EVEX_aaa = 0;
714 bool EncodeRC = false;
716 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_W)
719 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_L)
721 if (((TSFlags >> X86II::VEXShift) & X86II::EVEX_L2))
724 if (HasEVEX_K && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_Z))
727 if (((TSFlags >> X86II::VEXShift) & X86II::EVEX_B))
730 switch (TSFlags & X86II::OpPrefixMask) {
731 default: break; // VEX_PP already correct
732 case X86II::PD: VEX_PP = 0x1; break; // 66
733 case X86II::XS: VEX_PP = 0x2; break; // F3
734 case X86II::XD: VEX_PP = 0x3; break; // F2
737 switch (TSFlags & X86II::OpMapMask) {
738 default: llvm_unreachable("Invalid prefix!");
739 case X86II::TB: VEX_5M = 0x1; break; // 0F
740 case X86II::T8: VEX_5M = 0x2; break; // 0F 38
741 case X86II::TA: VEX_5M = 0x3; break; // 0F 3A
742 case X86II::XOP8: VEX_5M = 0x8; break;
743 case X86II::XOP9: VEX_5M = 0x9; break;
744 case X86II::XOPA: VEX_5M = 0xA; break;
747 // Classify VEX_B, VEX_4V, VEX_R, VEX_X
748 unsigned NumOps = Desc.getNumOperands();
749 unsigned CurOp = X86II::getOperandBias(Desc);
751 switch (TSFlags & X86II::FormMask) {
752 default: llvm_unreachable("Unexpected form in EmitVEXOpcodePrefix!");
755 case X86II::MRMDestMem: {
756 // MRMDestMem instructions forms:
757 // MemAddr, src1(ModR/M)
758 // MemAddr, src1(VEX_4V), src2(ModR/M)
759 // MemAddr, src1(ModR/M), imm8
761 if (X86II::isX86_64ExtendedReg(MI.getOperand(MemOperand +
762 X86::AddrBaseReg).getReg()))
764 if (X86II::isX86_64ExtendedReg(MI.getOperand(MemOperand +
765 X86::AddrIndexReg).getReg()))
767 if (X86II::is32ExtendedReg(MI.getOperand(MemOperand +
768 X86::AddrIndexReg).getReg()))
771 CurOp += X86::AddrNumOperands;
774 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
777 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
778 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
783 const MCOperand &MO = MI.getOperand(CurOp);
785 if (X86II::isX86_64ExtendedReg(MO.getReg()))
787 if (X86II::is32ExtendedReg(MO.getReg()))
792 case X86II::MRMSrcMem:
793 // MRMSrcMem instructions forms:
794 // src1(ModR/M), MemAddr
795 // src1(ModR/M), src2(VEX_4V), MemAddr
796 // src1(ModR/M), MemAddr, imm8
797 // src1(ModR/M), MemAddr, src2(VEX_I8IMM)
800 // dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
801 // dst(ModR/M.reg), src1(VEX_4V), src2(VEX_I8IMM), src3(ModR/M),
802 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
804 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
809 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
812 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
813 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
818 if (X86II::isX86_64ExtendedReg(
819 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
821 if (X86II::isX86_64ExtendedReg(
822 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
824 if (X86II::is32ExtendedReg(MI.getOperand(MemOperand +
825 X86::AddrIndexReg).getReg()))
829 // Instruction format for 4VOp3:
830 // src1(ModR/M), MemAddr, src3(VEX_4V)
831 // CurOp points to start of the MemoryOperand,
832 // it skips TIED_TO operands if exist, then increments past src1.
833 // CurOp + X86::AddrNumOperands will point to src3.
834 VEX_4V = getVEXRegisterEncoding(MI, CurOp+X86::AddrNumOperands);
836 case X86II::MRM0m: case X86II::MRM1m:
837 case X86II::MRM2m: case X86II::MRM3m:
838 case X86II::MRM4m: case X86II::MRM5m:
839 case X86II::MRM6m: case X86II::MRM7m: {
840 // MRM[0-9]m instructions forms:
842 // src1(VEX_4V), MemAddr
844 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
845 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
851 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
853 if (X86II::isX86_64ExtendedReg(
854 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
856 if (X86II::isX86_64ExtendedReg(
857 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
861 case X86II::MRMSrcReg:
862 // MRMSrcReg instructions forms:
863 // dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
864 // dst(ModR/M), src1(ModR/M)
865 // dst(ModR/M), src1(ModR/M), imm8
868 // dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
869 // dst(ModR/M.reg), src1(VEX_4V), src2(VEX_I8IMM), src3(ModR/M),
870 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
872 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
877 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
880 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
881 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
886 if (HasMemOp4) // Skip second register source (encoded in I8IMM)
889 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
891 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
895 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
898 unsigned RcOperand = NumOps-1;
899 assert(RcOperand >= CurOp);
900 EVEX_rc = MI.getOperand(RcOperand).getImm() & 0x3;
905 case X86II::MRMDestReg:
906 // MRMDestReg instructions forms:
907 // dst(ModR/M), src(ModR/M)
908 // dst(ModR/M), src(ModR/M), imm8
909 // dst(ModR/M), src1(VEX_4V), src2(ModR/M)
910 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
912 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
917 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
920 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
921 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
926 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
928 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
933 case X86II::MRM0r: case X86II::MRM1r:
934 case X86II::MRM2r: case X86II::MRM3r:
935 case X86II::MRM4r: case X86II::MRM5r:
936 case X86II::MRM6r: case X86II::MRM7r:
937 // MRM0r-MRM7r instructions forms:
938 // dst(VEX_4V), src(ModR/M), imm8
940 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
941 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
946 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
948 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
950 if (X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
955 if (Encoding == X86II::VEX || Encoding == X86II::XOP) {
956 // VEX opcode prefix can have 2 or 3 bytes
959 // +-----+ +--------------+ +-------------------+
960 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
961 // +-----+ +--------------+ +-------------------+
963 // +-----+ +-------------------+
964 // | C5h | | R | vvvv | L | pp |
965 // +-----+ +-------------------+
967 // XOP uses a similar prefix:
968 // +-----+ +--------------+ +-------------------+
969 // | 8Fh | | RXB | m-mmmm | | W | vvvv | L | pp |
970 // +-----+ +--------------+ +-------------------+
971 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
973 // Can we use the 2 byte VEX prefix?
974 if (Encoding == X86II::VEX && VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) {
975 EmitByte(0xC5, CurByte, OS);
976 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
981 EmitByte(Encoding == X86II::XOP ? 0x8F : 0xC4, CurByte, OS);
982 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
983 EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
985 assert(Encoding == X86II::EVEX && "unknown encoding!");
986 // EVEX opcode prefix can have 4 bytes
988 // +-----+ +--------------+ +-------------------+ +------------------------+
989 // | 62h | | RXBR' | 00mm | | W | vvvv | U | pp | | z | L'L | b | v' | aaa |
990 // +-----+ +--------------+ +-------------------+ +------------------------+
991 assert((VEX_5M & 0x3) == VEX_5M
992 && "More than 2 significant bits in VEX.m-mmmm fields for EVEX!");
996 EmitByte(0x62, CurByte, OS);
997 EmitByte((VEX_R << 7) |
1001 VEX_5M, CurByte, OS);
1002 EmitByte((VEX_W << 7) |
1005 VEX_PP, CurByte, OS);
1007 EmitByte((EVEX_z << 7) |
1011 EVEX_aaa, CurByte, OS);
1013 EmitByte((EVEX_z << 7) |
1018 EVEX_aaa, CurByte, OS);
1022 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
1023 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
1024 /// size, and 3) use of X86-64 extended registers.
1025 static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
1026 const MCInstrDesc &Desc) {
1028 if (TSFlags & X86II::REX_W)
1029 REX |= 1 << 3; // set REX.W
1031 if (MI.getNumOperands() == 0) return REX;
1033 unsigned NumOps = MI.getNumOperands();
1034 // FIXME: MCInst should explicitize the two-addrness.
1035 bool isTwoAddr = NumOps > 1 &&
1036 Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1;
1038 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
1039 unsigned i = isTwoAddr ? 1 : 0;
1040 for (; i != NumOps; ++i) {
1041 const MCOperand &MO = MI.getOperand(i);
1042 if (!MO.isReg()) continue;
1043 unsigned Reg = MO.getReg();
1044 if (!X86II::isX86_64NonExtLowByteReg(Reg)) continue;
1045 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
1046 // that returns non-zero.
1047 REX |= 0x40; // REX fixed encoding prefix
1051 switch (TSFlags & X86II::FormMask) {
1052 case X86II::MRMSrcReg:
1053 if (MI.getOperand(0).isReg() &&
1054 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
1055 REX |= 1 << 2; // set REX.R
1056 i = isTwoAddr ? 2 : 1;
1057 for (; i != NumOps; ++i) {
1058 const MCOperand &MO = MI.getOperand(i);
1059 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
1060 REX |= 1 << 0; // set REX.B
1063 case X86II::MRMSrcMem: {
1064 if (MI.getOperand(0).isReg() &&
1065 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
1066 REX |= 1 << 2; // set REX.R
1068 i = isTwoAddr ? 2 : 1;
1069 for (; i != NumOps; ++i) {
1070 const MCOperand &MO = MI.getOperand(i);
1072 if (X86II::isX86_64ExtendedReg(MO.getReg()))
1073 REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1)
1080 case X86II::MRM0m: case X86II::MRM1m:
1081 case X86II::MRM2m: case X86II::MRM3m:
1082 case X86II::MRM4m: case X86II::MRM5m:
1083 case X86II::MRM6m: case X86II::MRM7m:
1084 case X86II::MRMDestMem: {
1085 unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands);
1086 i = isTwoAddr ? 1 : 0;
1087 if (NumOps > e && MI.getOperand(e).isReg() &&
1088 X86II::isX86_64ExtendedReg(MI.getOperand(e).getReg()))
1089 REX |= 1 << 2; // set REX.R
1091 for (; i != e; ++i) {
1092 const MCOperand &MO = MI.getOperand(i);
1094 if (X86II::isX86_64ExtendedReg(MO.getReg()))
1095 REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1)
1102 if (MI.getOperand(0).isReg() &&
1103 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
1104 REX |= 1 << 0; // set REX.B
1105 i = isTwoAddr ? 2 : 1;
1106 for (unsigned e = NumOps; i != e; ++i) {
1107 const MCOperand &MO = MI.getOperand(i);
1108 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
1109 REX |= 1 << 2; // set REX.R
1116 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
1117 void X86MCCodeEmitter::EmitSegmentOverridePrefix(unsigned &CurByte,
1118 unsigned SegOperand,
1120 raw_ostream &OS) const {
1121 // Check for explicit segment override on memory operand.
1122 switch (MI.getOperand(SegOperand).getReg()) {
1123 default: llvm_unreachable("Unknown segment register!");
1125 case X86::CS: EmitByte(0x2E, CurByte, OS); break;
1126 case X86::SS: EmitByte(0x36, CurByte, OS); break;
1127 case X86::DS: EmitByte(0x3E, CurByte, OS); break;
1128 case X86::ES: EmitByte(0x26, CurByte, OS); break;
1129 case X86::FS: EmitByte(0x64, CurByte, OS); break;
1130 case X86::GS: EmitByte(0x65, CurByte, OS); break;
1134 /// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode.
1136 /// MemOperand is the operand # of the start of a memory operand if present. If
1137 /// Not present, it is -1.
1138 void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
1139 int MemOperand, const MCInst &MI,
1140 const MCInstrDesc &Desc,
1141 const MCSubtargetInfo &STI,
1142 raw_ostream &OS) const {
1144 // Emit the operand size opcode prefix as needed.
1145 unsigned char OpSize = (TSFlags & X86II::OpSizeMask) >> X86II::OpSizeShift;
1146 if (OpSize == (is16BitMode(STI) ? X86II::OpSize32 : X86II::OpSize16))
1147 EmitByte(0x66, CurByte, OS);
1149 switch (TSFlags & X86II::OpPrefixMask) {
1150 case X86II::PD: // 66
1151 EmitByte(0x66, CurByte, OS);
1153 case X86II::XS: // F3
1154 EmitByte(0xF3, CurByte, OS);
1156 case X86II::XD: // F2
1157 EmitByte(0xF2, CurByte, OS);
1161 // Handle REX prefix.
1162 // FIXME: Can this come before F2 etc to simplify emission?
1163 if (is64BitMode(STI)) {
1164 if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc))
1165 EmitByte(0x40 | REX, CurByte, OS);
1168 // 0x0F escape code must be emitted just before the opcode.
1169 switch (TSFlags & X86II::OpMapMask) {
1170 case X86II::TB: // Two-byte opcode map
1171 case X86II::T8: // 0F 38
1172 case X86II::TA: // 0F 3A
1173 EmitByte(0x0F, CurByte, OS);
1177 switch (TSFlags & X86II::OpMapMask) {
1178 case X86II::T8: // 0F 38
1179 EmitByte(0x38, CurByte, OS);
1181 case X86II::TA: // 0F 3A
1182 EmitByte(0x3A, CurByte, OS);
1187 void X86MCCodeEmitter::
1188 EncodeInstruction(const MCInst &MI, raw_ostream &OS,
1189 SmallVectorImpl<MCFixup> &Fixups,
1190 const MCSubtargetInfo &STI) const {
1191 unsigned Opcode = MI.getOpcode();
1192 const MCInstrDesc &Desc = MCII.get(Opcode);
1193 uint64_t TSFlags = Desc.TSFlags;
1195 // Pseudo instructions don't get encoded.
1196 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
1199 unsigned NumOps = Desc.getNumOperands();
1200 unsigned CurOp = X86II::getOperandBias(Desc);
1202 // Keep track of the current byte being emitted.
1203 unsigned CurByte = 0;
1205 // Encoding type for this instruction.
1206 unsigned char Encoding = (TSFlags & X86II::EncodingMask) >>
1207 X86II::EncodingShift;
1209 // It uses the VEX.VVVV field?
1210 bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
1211 bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
1212 bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4;
1213 const unsigned MemOp4_I8IMMOperand = 2;
1215 // It uses the EVEX.aaa field?
1216 bool HasEVEX_K = ((TSFlags >> X86II::VEXShift) & X86II::EVEX_K);
1217 bool HasEVEX_RC = ((TSFlags >> X86II::VEXShift) & X86II::EVEX_RC);
1219 // Determine where the memory operand starts, if present.
1220 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags, Opcode);
1221 if (MemoryOperand != -1) MemoryOperand += CurOp;
1223 // Emit the lock opcode prefix as needed.
1224 if (TSFlags & X86II::LOCK)
1225 EmitByte(0xF0, CurByte, OS);
1227 // Emit segment override opcode prefix as needed.
1228 if (MemoryOperand >= 0)
1229 EmitSegmentOverridePrefix(CurByte, MemoryOperand+X86::AddrSegmentReg,
1232 // Emit the repeat opcode prefix as needed.
1233 if (TSFlags & X86II::REP)
1234 EmitByte(0xF3, CurByte, OS);
1236 // Emit the address size opcode prefix as needed.
1237 bool need_address_override;
1238 // The AdSize prefix is only for 32-bit and 64-bit modes. Hm, perhaps we
1239 // should introduce an AdSize16 bit instead of having seven special cases?
1240 if ((!is16BitMode(STI) && TSFlags & X86II::AdSize) ||
1241 (is16BitMode(STI) && (MI.getOpcode() == X86::JECXZ_32 ||
1242 MI.getOpcode() == X86::MOV8o8a ||
1243 MI.getOpcode() == X86::MOV16o16a ||
1244 MI.getOpcode() == X86::MOV32o32a ||
1245 MI.getOpcode() == X86::MOV8ao8 ||
1246 MI.getOpcode() == X86::MOV16ao16 ||
1247 MI.getOpcode() == X86::MOV32ao32))) {
1248 need_address_override = true;
1249 } else if (MemoryOperand < 0) {
1250 need_address_override = false;
1251 } else if (is64BitMode(STI)) {
1252 assert(!Is16BitMemOperand(MI, MemoryOperand, STI));
1253 need_address_override = Is32BitMemOperand(MI, MemoryOperand);
1254 } else if (is32BitMode(STI)) {
1255 assert(!Is64BitMemOperand(MI, MemoryOperand));
1256 need_address_override = Is16BitMemOperand(MI, MemoryOperand, STI);
1258 assert(is16BitMode(STI));
1259 assert(!Is64BitMemOperand(MI, MemoryOperand));
1260 need_address_override = !Is16BitMemOperand(MI, MemoryOperand, STI);
1263 if (need_address_override)
1264 EmitByte(0x67, CurByte, OS);
1267 EmitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, STI, OS);
1269 EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
1271 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
1273 if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
1274 BaseOpcode = 0x0F; // Weird 3DNow! encoding.
1276 unsigned SrcRegNum = 0;
1277 switch (TSFlags & X86II::FormMask) {
1278 default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
1279 llvm_unreachable("Unknown FormMask value in X86MCCodeEmitter!");
1281 llvm_unreachable("Pseudo instruction shouldn't be emitted");
1282 case X86II::RawFrmDstSrc: {
1283 unsigned siReg = MI.getOperand(1).getReg();
1284 assert(((siReg == X86::SI && MI.getOperand(0).getReg() == X86::DI) ||
1285 (siReg == X86::ESI && MI.getOperand(0).getReg() == X86::EDI) ||
1286 (siReg == X86::RSI && MI.getOperand(0).getReg() == X86::RDI)) &&
1287 "SI and DI register sizes do not match");
1288 // Emit segment override opcode prefix as needed (not for %ds).
1289 if (MI.getOperand(2).getReg() != X86::DS)
1290 EmitSegmentOverridePrefix(CurByte, 2, MI, OS);
1291 // Emit AdSize prefix as needed.
1292 if ((!is32BitMode(STI) && siReg == X86::ESI) ||
1293 (is32BitMode(STI) && siReg == X86::SI))
1294 EmitByte(0x67, CurByte, OS);
1295 CurOp += 3; // Consume operands.
1296 EmitByte(BaseOpcode, CurByte, OS);
1299 case X86II::RawFrmSrc: {
1300 unsigned siReg = MI.getOperand(0).getReg();
1301 // Emit segment override opcode prefix as needed (not for %ds).
1302 if (MI.getOperand(1).getReg() != X86::DS)
1303 EmitSegmentOverridePrefix(CurByte, 1, MI, OS);
1304 // Emit AdSize prefix as needed.
1305 if ((!is32BitMode(STI) && siReg == X86::ESI) ||
1306 (is32BitMode(STI) && siReg == X86::SI))
1307 EmitByte(0x67, CurByte, OS);
1308 CurOp += 2; // Consume operands.
1309 EmitByte(BaseOpcode, CurByte, OS);
1312 case X86II::RawFrmDst: {
1313 unsigned siReg = MI.getOperand(0).getReg();
1314 // Emit AdSize prefix as needed.
1315 if ((!is32BitMode(STI) && siReg == X86::EDI) ||
1316 (is32BitMode(STI) && siReg == X86::DI))
1317 EmitByte(0x67, CurByte, OS);
1318 ++CurOp; // Consume operand.
1319 EmitByte(BaseOpcode, CurByte, OS);
1323 EmitByte(BaseOpcode, CurByte, OS);
1325 case X86II::RawFrmMemOffs:
1326 // Emit segment override opcode prefix as needed.
1327 EmitSegmentOverridePrefix(CurByte, 1, MI, OS);
1328 EmitByte(BaseOpcode, CurByte, OS);
1329 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1330 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1331 CurByte, OS, Fixups);
1332 ++CurOp; // skip segment operand
1334 case X86II::RawFrmImm8:
1335 EmitByte(BaseOpcode, CurByte, OS);
1336 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1337 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1338 CurByte, OS, Fixups);
1339 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 1, FK_Data_1, CurByte,
1342 case X86II::RawFrmImm16:
1343 EmitByte(BaseOpcode, CurByte, OS);
1344 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1345 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1346 CurByte, OS, Fixups);
1347 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 2, FK_Data_2, CurByte,
1351 case X86II::AddRegFrm:
1352 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
1355 case X86II::MRMDestReg:
1356 EmitByte(BaseOpcode, CurByte, OS);
1357 SrcRegNum = CurOp + 1;
1359 if (HasEVEX_K) // Skip writemask
1362 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1365 EmitRegModRMByte(MI.getOperand(CurOp),
1366 GetX86RegNum(MI.getOperand(SrcRegNum)), CurByte, OS);
1367 CurOp = SrcRegNum + 1;
1370 case X86II::MRMDestMem:
1371 EmitByte(BaseOpcode, CurByte, OS);
1372 SrcRegNum = CurOp + X86::AddrNumOperands;
1374 if (HasEVEX_K) // Skip writemask
1377 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1380 EmitMemModRMByte(MI, CurOp,
1381 GetX86RegNum(MI.getOperand(SrcRegNum)),
1382 TSFlags, CurByte, OS, Fixups, STI);
1383 CurOp = SrcRegNum + 1;
1386 case X86II::MRMSrcReg:
1387 EmitByte(BaseOpcode, CurByte, OS);
1388 SrcRegNum = CurOp + 1;
1390 if (HasEVEX_K) // Skip writemask
1393 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1396 if (HasMemOp4) // Skip 2nd src (which is encoded in I8IMM)
1399 EmitRegModRMByte(MI.getOperand(SrcRegNum),
1400 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1402 // 2 operands skipped with HasMemOp4, compensate accordingly
1403 CurOp = HasMemOp4 ? SrcRegNum : SrcRegNum + 1;
1406 // do not count the rounding control operand
1411 case X86II::MRMSrcMem: {
1412 int AddrOperands = X86::AddrNumOperands;
1413 unsigned FirstMemOp = CurOp+1;
1415 if (HasEVEX_K) { // Skip writemask
1422 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
1424 if (HasMemOp4) // Skip second register source (encoded in I8IMM)
1427 EmitByte(BaseOpcode, CurByte, OS);
1429 EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1430 TSFlags, CurByte, OS, Fixups, STI);
1431 CurOp += AddrOperands + 1;
1438 case X86II::MRM0r: case X86II::MRM1r:
1439 case X86II::MRM2r: case X86II::MRM3r:
1440 case X86II::MRM4r: case X86II::MRM5r:
1441 case X86II::MRM6r: case X86II::MRM7r: {
1442 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1444 if (HasEVEX_K) // Skip writemask
1446 EmitByte(BaseOpcode, CurByte, OS);
1447 uint64_t Form = TSFlags & X86II::FormMask;
1448 EmitRegModRMByte(MI.getOperand(CurOp++),
1449 (Form == X86II::MRMXr) ? 0 : Form-X86II::MRM0r,
1455 case X86II::MRM0m: case X86II::MRM1m:
1456 case X86II::MRM2m: case X86II::MRM3m:
1457 case X86II::MRM4m: case X86II::MRM5m:
1458 case X86II::MRM6m: case X86II::MRM7m: {
1459 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1461 if (HasEVEX_K) // Skip writemask
1463 EmitByte(BaseOpcode, CurByte, OS);
1464 uint64_t Form = TSFlags & X86II::FormMask;
1465 EmitMemModRMByte(MI, CurOp, (Form == X86II::MRMXm) ? 0 : Form-X86II::MRM0m,
1466 TSFlags, CurByte, OS, Fixups, STI);
1467 CurOp += X86::AddrNumOperands;
1470 case X86II::MRM_C0: case X86II::MRM_C1: case X86II::MRM_C2:
1471 case X86II::MRM_C3: case X86II::MRM_C4: case X86II::MRM_C8:
1472 case X86II::MRM_C9: case X86II::MRM_CA: case X86II::MRM_CB:
1473 case X86II::MRM_D0: case X86II::MRM_D1: case X86II::MRM_D4:
1474 case X86II::MRM_D5: case X86II::MRM_D6: case X86II::MRM_D8:
1475 case X86II::MRM_D9: case X86II::MRM_DA: case X86II::MRM_DB:
1476 case X86II::MRM_DC: case X86II::MRM_DD: case X86II::MRM_DE:
1477 case X86II::MRM_DF: case X86II::MRM_E0: case X86II::MRM_E1:
1478 case X86II::MRM_E2: case X86II::MRM_E3: case X86II::MRM_E4:
1479 case X86II::MRM_E5: case X86II::MRM_E8: case X86II::MRM_E9:
1480 case X86II::MRM_EA: case X86II::MRM_EB: case X86II::MRM_EC:
1481 case X86II::MRM_ED: case X86II::MRM_EE: case X86II::MRM_F0:
1482 case X86II::MRM_F1: case X86II::MRM_F2: case X86II::MRM_F3:
1483 case X86II::MRM_F4: case X86II::MRM_F5: case X86II::MRM_F6:
1484 case X86II::MRM_F7: case X86II::MRM_F8: case X86II::MRM_F9:
1485 case X86II::MRM_FA: case X86II::MRM_FB: case X86II::MRM_FC:
1486 case X86II::MRM_FD: case X86II::MRM_FE: case X86II::MRM_FF:
1487 EmitByte(BaseOpcode, CurByte, OS);
1490 switch (TSFlags & X86II::FormMask) {
1491 default: llvm_unreachable("Invalid Form");
1492 case X86II::MRM_C0: MRM = 0xC0; break;
1493 case X86II::MRM_C1: MRM = 0xC1; break;
1494 case X86II::MRM_C2: MRM = 0xC2; break;
1495 case X86II::MRM_C3: MRM = 0xC3; break;
1496 case X86II::MRM_C4: MRM = 0xC4; break;
1497 case X86II::MRM_C8: MRM = 0xC8; break;
1498 case X86II::MRM_C9: MRM = 0xC9; break;
1499 case X86II::MRM_CA: MRM = 0xCA; break;
1500 case X86II::MRM_CB: MRM = 0xCB; break;
1501 case X86II::MRM_D0: MRM = 0xD0; break;
1502 case X86II::MRM_D1: MRM = 0xD1; break;
1503 case X86II::MRM_D4: MRM = 0xD4; break;
1504 case X86II::MRM_D5: MRM = 0xD5; break;
1505 case X86II::MRM_D6: MRM = 0xD6; break;
1506 case X86II::MRM_D8: MRM = 0xD8; break;
1507 case X86II::MRM_D9: MRM = 0xD9; break;
1508 case X86II::MRM_DA: MRM = 0xDA; break;
1509 case X86II::MRM_DB: MRM = 0xDB; break;
1510 case X86II::MRM_DC: MRM = 0xDC; break;
1511 case X86II::MRM_DD: MRM = 0xDD; break;
1512 case X86II::MRM_DE: MRM = 0xDE; break;
1513 case X86II::MRM_DF: MRM = 0xDF; break;
1514 case X86II::MRM_E0: MRM = 0xE0; break;
1515 case X86II::MRM_E1: MRM = 0xE1; break;
1516 case X86II::MRM_E2: MRM = 0xE2; break;
1517 case X86II::MRM_E3: MRM = 0xE3; break;
1518 case X86II::MRM_E4: MRM = 0xE4; break;
1519 case X86II::MRM_E5: MRM = 0xE5; break;
1520 case X86II::MRM_E8: MRM = 0xE8; break;
1521 case X86II::MRM_E9: MRM = 0xE9; break;
1522 case X86II::MRM_EA: MRM = 0xEA; break;
1523 case X86II::MRM_EB: MRM = 0xEB; break;
1524 case X86II::MRM_EC: MRM = 0xEC; break;
1525 case X86II::MRM_ED: MRM = 0xED; break;
1526 case X86II::MRM_EE: MRM = 0xEE; break;
1527 case X86II::MRM_F0: MRM = 0xF0; break;
1528 case X86II::MRM_F1: MRM = 0xF1; break;
1529 case X86II::MRM_F2: MRM = 0xF2; break;
1530 case X86II::MRM_F3: MRM = 0xF3; break;
1531 case X86II::MRM_F4: MRM = 0xF4; break;
1532 case X86II::MRM_F5: MRM = 0xF5; break;
1533 case X86II::MRM_F6: MRM = 0xF6; break;
1534 case X86II::MRM_F7: MRM = 0xF7; break;
1535 case X86II::MRM_F8: MRM = 0xF8; break;
1536 case X86II::MRM_F9: MRM = 0xF9; break;
1537 case X86II::MRM_FA: MRM = 0xFA; break;
1538 case X86II::MRM_FB: MRM = 0xFB; break;
1539 case X86II::MRM_FC: MRM = 0xFC; break;
1540 case X86II::MRM_FD: MRM = 0xFD; break;
1541 case X86II::MRM_FE: MRM = 0xFE; break;
1542 case X86II::MRM_FF: MRM = 0xFF; break;
1544 EmitByte(MRM, CurByte, OS);
1548 // If there is a remaining operand, it must be a trailing immediate. Emit it
1549 // according to the right size for the instruction. Some instructions
1550 // (SSE4a extrq and insertq) have two trailing immediates.
1551 while (CurOp != NumOps && NumOps - CurOp <= 2) {
1552 // The last source register of a 4 operand instruction in AVX is encoded
1553 // in bits[7:4] of a immediate byte.
1554 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) {
1555 const MCOperand &MO = MI.getOperand(HasMemOp4 ? MemOp4_I8IMMOperand
1558 unsigned RegNum = GetX86RegNum(MO) << 4;
1559 if (X86II::isX86_64ExtendedReg(MO.getReg()))
1561 // If there is an additional 5th operand it must be an immediate, which
1562 // is encoded in bits[3:0]
1563 if (CurOp != NumOps) {
1564 const MCOperand &MIMM = MI.getOperand(CurOp++);
1566 unsigned Val = MIMM.getImm();
1567 assert(Val < 16 && "Immediate operand value out of range");
1571 EmitImmediate(MCOperand::CreateImm(RegNum), MI.getLoc(), 1, FK_Data_1,
1572 CurByte, OS, Fixups);
1574 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1575 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1576 CurByte, OS, Fixups);
1580 if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
1581 EmitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS);
1585 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
1586 errs() << "Cannot encode all operands of: ";