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/X86BaseInfo.h"
15 #include "MCTargetDesc/X86FixupKinds.h"
16 #include "MCTargetDesc/X86MCTargetDesc.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/MC/MCCodeEmitter.h"
19 #include "llvm/MC/MCContext.h"
20 #include "llvm/MC/MCExpr.h"
21 #include "llvm/MC/MCFixup.h"
22 #include "llvm/MC/MCInst.h"
23 #include "llvm/MC/MCInstrDesc.h"
24 #include "llvm/MC/MCInstrInfo.h"
25 #include "llvm/MC/MCRegisterInfo.h"
26 #include "llvm/MC/MCSubtargetInfo.h"
27 #include "llvm/MC/MCSymbol.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/raw_ostream.h"
36 #define DEBUG_TYPE "mccodeemitter"
40 class X86MCCodeEmitter : public MCCodeEmitter {
41 const MCInstrInfo &MCII;
45 X86MCCodeEmitter(const MCInstrInfo &mcii, MCContext &ctx)
46 : MCII(mcii), Ctx(ctx) {
48 X86MCCodeEmitter(const X86MCCodeEmitter &) = delete;
49 X86MCCodeEmitter &operator=(const X86MCCodeEmitter &) = delete;
50 ~X86MCCodeEmitter() override = default;
52 bool is64BitMode(const MCSubtargetInfo &STI) const {
53 return STI.getFeatureBits()[X86::Mode64Bit];
56 bool is32BitMode(const MCSubtargetInfo &STI) const {
57 return STI.getFeatureBits()[X86::Mode32Bit];
60 bool is16BitMode(const MCSubtargetInfo &STI) const {
61 return STI.getFeatureBits()[X86::Mode16Bit];
64 /// Is16BitMemOperand - Return true if the specified instruction has
65 /// a 16-bit memory operand. Op specifies the operand # of the memoperand.
66 bool Is16BitMemOperand(const MCInst &MI, unsigned Op,
67 const MCSubtargetInfo &STI) const {
68 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
69 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
70 const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
72 if (is16BitMode(STI) && BaseReg.getReg() == 0 &&
73 Disp.isImm() && Disp.getImm() < 0x10000)
75 if ((BaseReg.getReg() != 0 &&
76 X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) ||
77 (IndexReg.getReg() != 0 &&
78 X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg())))
83 unsigned GetX86RegNum(const MCOperand &MO) const {
84 return Ctx.getRegisterInfo()->getEncodingValue(MO.getReg()) & 0x7;
87 unsigned getX86RegEncoding(const MCInst &MI, unsigned OpNum) const {
88 return Ctx.getRegisterInfo()->getEncodingValue(
89 MI.getOperand(OpNum).getReg());
92 // Does this register require a bit to be set in REX prefix.
93 bool isREXExtendedReg(const MCInst &MI, unsigned OpNum) const {
94 return (getX86RegEncoding(MI, OpNum) >> 3) & 1;
97 void EmitByte(uint8_t C, unsigned &CurByte, raw_ostream &OS) const {
102 void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
103 raw_ostream &OS) const {
104 // Output the constant in little endian byte order.
105 for (unsigned i = 0; i != Size; ++i) {
106 EmitByte(Val & 255, CurByte, OS);
111 void EmitImmediate(const MCOperand &Disp, SMLoc Loc,
112 unsigned ImmSize, MCFixupKind FixupKind,
113 unsigned &CurByte, raw_ostream &OS,
114 SmallVectorImpl<MCFixup> &Fixups,
115 int ImmOffset = 0) const;
117 static uint8_t ModRMByte(unsigned Mod, unsigned RegOpcode, unsigned RM) {
118 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
119 return RM | (RegOpcode << 3) | (Mod << 6);
122 void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
123 unsigned &CurByte, raw_ostream &OS) const {
124 EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS);
127 void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base,
128 unsigned &CurByte, raw_ostream &OS) const {
129 // SIB byte is in the same format as the ModRMByte.
130 EmitByte(ModRMByte(SS, Index, Base), CurByte, OS);
133 void emitMemModRMByte(const MCInst &MI, unsigned Op, unsigned RegOpcodeField,
134 uint64_t TSFlags, bool Rex, unsigned &CurByte,
135 raw_ostream &OS, SmallVectorImpl<MCFixup> &Fixups,
136 const MCSubtargetInfo &STI) const;
138 void encodeInstruction(const MCInst &MI, raw_ostream &OS,
139 SmallVectorImpl<MCFixup> &Fixups,
140 const MCSubtargetInfo &STI) const override;
142 void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
143 const MCInst &MI, const MCInstrDesc &Desc,
144 raw_ostream &OS) const;
146 void EmitSegmentOverridePrefix(unsigned &CurByte, unsigned SegOperand,
147 const MCInst &MI, raw_ostream &OS) const;
149 bool emitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
150 const MCInst &MI, const MCInstrDesc &Desc,
151 const MCSubtargetInfo &STI, raw_ostream &OS) const;
153 uint8_t DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
154 int MemOperand, const MCInstrDesc &Desc) const;
156 bool isPCRel32Branch(const MCInst &MI) const;
159 } // end anonymous namespace
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 == (int8_t)Value;
167 /// isCDisp8 - Return true if this signed displacement fits in a 8-bit
168 /// compressed dispacement field.
169 static bool isCDisp8(uint64_t TSFlags, int Value, int& CValue) {
170 assert(((TSFlags & X86II::EncodingMask) == X86II::EVEX) &&
171 "Compressed 8-bit displacement is only valid for EVEX inst.");
174 (TSFlags & X86II::CD8_Scale_Mask) >> X86II::CD8_Scale_Shift;
175 if (CD8_Scale == 0) {
177 return isDisp8(Value);
180 unsigned Mask = CD8_Scale - 1;
181 assert((CD8_Scale & Mask) == 0 && "Invalid memory object size.");
182 if (Value & Mask) // Unaligned offset
184 Value /= (int)CD8_Scale;
185 bool Ret = (Value == (int8_t)Value);
192 /// getImmFixupKind - Return the appropriate fixup kind to use for an immediate
193 /// in an instruction with the specified TSFlags.
194 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
195 unsigned Size = X86II::getSizeOfImm(TSFlags);
196 bool isPCRel = X86II::isImmPCRel(TSFlags);
198 if (X86II::isImmSigned(TSFlags)) {
200 default: llvm_unreachable("Unsupported signed fixup size!");
201 case 4: return MCFixupKind(X86::reloc_signed_4byte);
204 return MCFixup::getKindForSize(Size, isPCRel);
207 /// Is32BitMemOperand - Return true if the specified instruction has
208 /// a 32-bit memory operand. Op specifies the operand # of the memoperand.
209 static bool Is32BitMemOperand(const MCInst &MI, unsigned Op) {
210 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
211 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
213 if ((BaseReg.getReg() != 0 &&
214 X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) ||
215 (IndexReg.getReg() != 0 &&
216 X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg())))
218 if (BaseReg.getReg() == X86::EIP) {
219 assert(IndexReg.getReg() == 0 && "Invalid eip-based address.");
225 /// Is64BitMemOperand - Return true if the specified instruction has
226 /// a 64-bit memory operand. Op specifies the operand # of the memoperand.
228 static bool Is64BitMemOperand(const MCInst &MI, unsigned Op) {
229 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
230 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
232 if ((BaseReg.getReg() != 0 &&
233 X86MCRegisterClasses[X86::GR64RegClassID].contains(BaseReg.getReg())) ||
234 (IndexReg.getReg() != 0 &&
235 X86MCRegisterClasses[X86::GR64RegClassID].contains(IndexReg.getReg())))
241 /// StartsWithGlobalOffsetTable - Check if this expression starts with
242 /// _GLOBAL_OFFSET_TABLE_ and if it is of the form
243 /// _GLOBAL_OFFSET_TABLE_-symbol. This is needed to support PIC on ELF
244 /// i386 as _GLOBAL_OFFSET_TABLE_ is magical. We check only simple case that
245 /// are know to be used: _GLOBAL_OFFSET_TABLE_ by itself or at the start
246 /// of a binary expression.
247 enum GlobalOffsetTableExprKind {
252 static GlobalOffsetTableExprKind
253 StartsWithGlobalOffsetTable(const MCExpr *Expr) {
254 const MCExpr *RHS = nullptr;
255 if (Expr->getKind() == MCExpr::Binary) {
256 const MCBinaryExpr *BE = static_cast<const MCBinaryExpr *>(Expr);
261 if (Expr->getKind() != MCExpr::SymbolRef)
264 const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
265 const MCSymbol &S = Ref->getSymbol();
266 if (S.getName() != "_GLOBAL_OFFSET_TABLE_")
268 if (RHS && RHS->getKind() == MCExpr::SymbolRef)
273 static bool HasSecRelSymbolRef(const MCExpr *Expr) {
274 if (Expr->getKind() == MCExpr::SymbolRef) {
275 const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
276 return Ref->getKind() == MCSymbolRefExpr::VK_SECREL;
281 bool X86MCCodeEmitter::isPCRel32Branch(const MCInst &MI) const {
282 unsigned Opcode = MI.getOpcode();
283 const MCInstrDesc &Desc = MCII.get(Opcode);
284 if ((Opcode != X86::CALL64pcrel32 && Opcode != X86::JMP_4) ||
285 getImmFixupKind(Desc.TSFlags) != FK_PCRel_4)
288 unsigned CurOp = X86II::getOperandBias(Desc);
289 const MCOperand &Op = MI.getOperand(CurOp);
293 const MCSymbolRefExpr *Ref = dyn_cast<MCSymbolRefExpr>(Op.getExpr());
294 return Ref && Ref->getKind() == MCSymbolRefExpr::VK_None;
297 void X86MCCodeEmitter::
298 EmitImmediate(const MCOperand &DispOp, SMLoc Loc, unsigned Size,
299 MCFixupKind FixupKind, unsigned &CurByte, raw_ostream &OS,
300 SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
301 const MCExpr *Expr = nullptr;
302 if (DispOp.isImm()) {
303 // If this is a simple integer displacement that doesn't require a
304 // relocation, emit it now.
305 if (FixupKind != FK_PCRel_1 &&
306 FixupKind != FK_PCRel_2 &&
307 FixupKind != FK_PCRel_4) {
308 EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
311 Expr = MCConstantExpr::create(DispOp.getImm(), Ctx);
313 Expr = DispOp.getExpr();
316 // If we have an immoffset, add it to the expression.
317 if ((FixupKind == FK_Data_4 ||
318 FixupKind == FK_Data_8 ||
319 FixupKind == MCFixupKind(X86::reloc_signed_4byte))) {
320 GlobalOffsetTableExprKind Kind = StartsWithGlobalOffsetTable(Expr);
321 if (Kind != GOT_None) {
322 assert(ImmOffset == 0);
325 FixupKind = MCFixupKind(X86::reloc_global_offset_table8);
328 FixupKind = MCFixupKind(X86::reloc_global_offset_table);
331 if (Kind == GOT_Normal)
333 } else if (Expr->getKind() == MCExpr::SymbolRef) {
334 if (HasSecRelSymbolRef(Expr)) {
335 FixupKind = MCFixupKind(FK_SecRel_4);
337 } else if (Expr->getKind() == MCExpr::Binary) {
338 const MCBinaryExpr *Bin = static_cast<const MCBinaryExpr*>(Expr);
339 if (HasSecRelSymbolRef(Bin->getLHS())
340 || HasSecRelSymbolRef(Bin->getRHS())) {
341 FixupKind = MCFixupKind(FK_SecRel_4);
346 // If the fixup is pc-relative, we need to bias the value to be relative to
347 // the start of the field, not the end of the field.
348 if (FixupKind == FK_PCRel_4 ||
349 FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
350 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load) ||
351 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_relax) ||
352 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_relax_rex) ||
353 FixupKind == MCFixupKind(X86::reloc_branch_4byte_pcrel))
355 if (FixupKind == FK_PCRel_2)
357 if (FixupKind == FK_PCRel_1)
361 Expr = MCBinaryExpr::createAdd(Expr, MCConstantExpr::create(ImmOffset, Ctx),
364 // Emit a symbolic constant as a fixup and 4 zeros.
365 Fixups.push_back(MCFixup::create(CurByte, Expr, FixupKind, Loc));
366 EmitConstant(0, Size, CurByte, OS);
369 void X86MCCodeEmitter::emitMemModRMByte(const MCInst &MI, unsigned Op,
370 unsigned RegOpcodeField,
371 uint64_t TSFlags, bool Rex,
372 unsigned &CurByte, raw_ostream &OS,
373 SmallVectorImpl<MCFixup> &Fixups,
374 const MCSubtargetInfo &STI) const {
375 const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
376 const MCOperand &Base = MI.getOperand(Op+X86::AddrBaseReg);
377 const MCOperand &Scale = MI.getOperand(Op+X86::AddrScaleAmt);
378 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
379 unsigned BaseReg = Base.getReg();
380 bool HasEVEX = (TSFlags & X86II::EncodingMask) == X86II::EVEX;
382 // Handle %rip relative addressing.
383 if (BaseReg == X86::RIP ||
384 BaseReg == X86::EIP) { // [disp32+rIP] in X86-64 mode
385 assert(is64BitMode(STI) && "Rip-relative addressing requires 64-bit mode");
386 assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
387 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
389 unsigned Opcode = MI.getOpcode();
390 // movq loads are handled with a special relocation form which allows the
391 // linker to eliminate some loads for GOT references which end up in the
392 // same linkage unit.
393 unsigned FixupKind = [=]() {
396 return X86::reloc_riprel_4byte;
399 return X86::reloc_riprel_4byte_movq_load;
411 return Rex ? X86::reloc_riprel_4byte_relax_rex
412 : X86::reloc_riprel_4byte_relax;
416 // rip-relative addressing is actually relative to the *next* instruction.
417 // Since an immediate can follow the mod/rm byte for an instruction, this
418 // means that we need to bias the displacement field of the instruction with
419 // the size of the immediate field. If we have this case, add it into the
420 // expression to emit.
421 // Note: rip-relative addressing using immediate displacement values should
422 // not be adjusted, assuming it was the user's intent.
423 int ImmSize = !Disp.isImm() && X86II::hasImm(TSFlags)
424 ? X86II::getSizeOfImm(TSFlags)
427 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind),
428 CurByte, OS, Fixups, -ImmSize);
432 unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
434 // 16-bit addressing forms of the ModR/M byte have a different encoding for
435 // the R/M field and are far more limited in which registers can be used.
436 if (Is16BitMemOperand(MI, Op, STI)) {
438 // For 32-bit addressing, the row and column values in Table 2-2 are
439 // basically the same. It's AX/CX/DX/BX/SP/BP/SI/DI in that order, with
440 // some special cases. And GetX86RegNum reflects that numbering.
441 // For 16-bit addressing it's more fun, as shown in the SDM Vol 2A,
442 // Table 2-1 "16-Bit Addressing Forms with the ModR/M byte". We can only
443 // use SI/DI/BP/BX, which have "row" values 4-7 in no particular order,
444 // while values 0-3 indicate the allowed combinations (base+index) of
445 // those: 0 for BX+SI, 1 for BX+DI, 2 for BP+SI, 3 for BP+DI.
447 // R16Table[] is a lookup from the normal RegNo, to the row values from
448 // Table 2-1 for 16-bit addressing modes. Where zero means disallowed.
449 static const unsigned R16Table[] = { 0, 0, 0, 7, 0, 6, 4, 5 };
450 unsigned RMfield = R16Table[BaseRegNo];
452 assert(RMfield && "invalid 16-bit base register");
454 if (IndexReg.getReg()) {
455 unsigned IndexReg16 = R16Table[GetX86RegNum(IndexReg)];
457 assert(IndexReg16 && "invalid 16-bit index register");
458 // We must have one of SI/DI (4,5), and one of BP/BX (6,7).
459 assert(((IndexReg16 ^ RMfield) & 2) &&
460 "invalid 16-bit base/index register combination");
461 assert(Scale.getImm() == 1 &&
462 "invalid scale for 16-bit memory reference");
464 // Allow base/index to appear in either order (although GAS doesn't).
466 RMfield = (RMfield & 1) | ((7 - IndexReg16) << 1);
468 RMfield = (IndexReg16 & 1) | ((7 - RMfield) << 1);
471 if (Disp.isImm() && isDisp8(Disp.getImm())) {
472 if (Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
473 // There is no displacement; just the register.
474 EmitByte(ModRMByte(0, RegOpcodeField, RMfield), CurByte, OS);
477 // Use the [REG]+disp8 form, including for [BP] which cannot be encoded.
478 EmitByte(ModRMByte(1, RegOpcodeField, RMfield), CurByte, OS);
479 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
482 // This is the [REG]+disp16 case.
483 EmitByte(ModRMByte(2, RegOpcodeField, RMfield), CurByte, OS);
485 // There is no BaseReg; this is the plain [disp16] case.
486 EmitByte(ModRMByte(0, RegOpcodeField, 6), CurByte, OS);
489 // Emit 16-bit displacement for plain disp16 or [REG]+disp16 cases.
490 EmitImmediate(Disp, MI.getLoc(), 2, FK_Data_2, CurByte, OS, Fixups);
494 // Determine whether a SIB byte is needed.
495 // If no BaseReg, issue a RIP relative instruction only if the MCE can
496 // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
497 // 2-7) and absolute references.
499 if (// The SIB byte must be used if there is an index register.
500 IndexReg.getReg() == 0 &&
501 // The SIB byte must be used if the base is ESP/RSP/R12, all of which
502 // encode to an R/M value of 4, which indicates that a SIB byte is
504 BaseRegNo != N86::ESP &&
505 // If there is no base register and we're in 64-bit mode, we need a SIB
506 // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
507 (!is64BitMode(STI) || BaseReg != 0)) {
509 if (BaseReg == 0) { // [disp32] in X86-32 mode
510 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
511 EmitImmediate(Disp, MI.getLoc(), 4, FK_Data_4, CurByte, OS, Fixups);
515 // If the base is not EBP/ESP and there is no displacement, use simple
516 // indirect register encoding, this handles addresses like [EAX]. The
517 // encoding for [EBP] with no displacement means [disp32] so we handle it
518 // by emitting a displacement of 0 below.
519 if (Disp.isImm() && Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
520 EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
524 // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
526 if (!HasEVEX && isDisp8(Disp.getImm())) {
527 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
528 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
531 // Try EVEX compressed 8-bit displacement first; if failed, fall back to
532 // 32-bit displacement.
534 if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
535 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
536 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups,
537 CDisp8 - Disp.getImm());
542 // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
543 EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
544 unsigned Opcode = MI.getOpcode();
545 unsigned FixupKind = Opcode == X86::MOV32rm ? X86::reloc_signed_4byte_relax
546 : X86::reloc_signed_4byte;
547 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind), CurByte, OS,
552 // We need a SIB byte, so start by outputting the ModR/M byte first
553 assert(IndexReg.getReg() != X86::ESP &&
554 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
556 bool ForceDisp32 = false;
557 bool ForceDisp8 = false;
561 // If there is no base register, we emit the special case SIB byte with
562 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
563 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
565 } else if (!Disp.isImm()) {
566 // Emit the normal disp32 encoding.
567 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
569 } else if (Disp.getImm() == 0 &&
570 // Base reg can't be anything that ends up with '5' as the base
571 // reg, it is the magic [*] nomenclature that indicates no base.
572 BaseRegNo != N86::EBP) {
573 // Emit no displacement ModR/M byte
574 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
575 } else if (!HasEVEX && isDisp8(Disp.getImm())) {
576 // Emit the disp8 encoding.
577 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
578 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
579 } else if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
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 ImmOffset = CDisp8 - Disp.getImm();
585 // Emit the normal disp32 encoding.
586 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
589 // Calculate what the SS field value should be...
590 static const unsigned SSTable[] = { ~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3 };
591 unsigned SS = SSTable[Scale.getImm()];
594 // Handle the SIB byte for the case where there is no base, see Intel
595 // Manual 2A, table 2-7. The displacement has already been output.
597 if (IndexReg.getReg())
598 IndexRegNo = GetX86RegNum(IndexReg);
599 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
601 EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
604 if (IndexReg.getReg())
605 IndexRegNo = GetX86RegNum(IndexReg);
607 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
608 EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
611 // Do we need to output a displacement?
613 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups, ImmOffset);
614 else if (ForceDisp32 || Disp.getImm() != 0)
615 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(X86::reloc_signed_4byte),
616 CurByte, OS, Fixups);
619 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
621 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
622 int MemOperand, const MCInst &MI,
623 const MCInstrDesc &Desc,
624 raw_ostream &OS) const {
625 assert(!(TSFlags & X86II::LOCK) && "Can't have LOCK VEX.");
627 uint64_t Encoding = TSFlags & X86II::EncodingMask;
628 bool HasEVEX_K = TSFlags & X86II::EVEX_K;
629 bool HasVEX_4V = TSFlags & X86II::VEX_4V;
630 bool HasEVEX_RC = TSFlags & X86II::EVEX_RC;
632 // VEX_R: opcode externsion equivalent to REX.R in
633 // 1's complement (inverted) form
635 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
636 // 0: Same as REX_R=1 (64 bit mode only)
639 uint8_t EVEX_R2 = 0x1;
641 // VEX_X: equivalent to REX.X, only used when a
642 // register is used for index in SIB Byte.
644 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
645 // 0: Same as REX.X=1 (64-bit mode only)
650 // 1: Same as REX_B=0 (ignored in 32-bit mode)
651 // 0: Same as REX_B=1 (64 bit mode only)
655 // VEX_W: opcode specific (use like REX.W, or used for
656 // opcode extension, or ignored, depending on the opcode byte)
657 uint8_t VEX_W = (TSFlags & X86II::VEX_W) ? 1 : 0;
659 // VEX_5M (VEX m-mmmmm field):
661 // 0b00000: Reserved for future use
662 // 0b00001: implied 0F leading opcode
663 // 0b00010: implied 0F 38 leading opcode bytes
664 // 0b00011: implied 0F 3A leading opcode bytes
665 // 0b00100-0b11111: Reserved for future use
666 // 0b01000: XOP map select - 08h instructions with imm byte
667 // 0b01001: XOP map select - 09h instructions with no imm byte
668 // 0b01010: XOP map select - 0Ah instructions with imm dword
670 switch (TSFlags & X86II::OpMapMask) {
671 default: llvm_unreachable("Invalid prefix!");
672 case X86II::TB: VEX_5M = 0x1; break; // 0F
673 case X86II::T8: VEX_5M = 0x2; break; // 0F 38
674 case X86II::TA: VEX_5M = 0x3; break; // 0F 3A
675 case X86II::XOP8: VEX_5M = 0x8; break;
676 case X86II::XOP9: VEX_5M = 0x9; break;
677 case X86II::XOPA: VEX_5M = 0xA; break;
680 // VEX_4V (VEX vvvv field): a register specifier
681 // (in 1's complement form) or 1111 if unused.
682 uint8_t VEX_4V = 0xf;
683 uint8_t EVEX_V2 = 0x1;
685 // EVEX_L2/VEX_L (Vector Length):
688 // 0 0: scalar or 128-bit vector
689 // 0 1: 256-bit vector
690 // 1 0: 512-bit vector
692 uint8_t VEX_L = (TSFlags & X86II::VEX_L) ? 1 : 0;
693 uint8_t EVEX_L2 = (TSFlags & X86II::EVEX_L2) ? 1 : 0;
695 // VEX_PP: opcode extension providing equivalent
696 // functionality of a SIMD prefix
704 switch (TSFlags & X86II::OpPrefixMask) {
705 default: llvm_unreachable("Invalid op prefix!");
706 case X86II::PS: VEX_PP = 0x0; break; // none
707 case X86II::PD: VEX_PP = 0x1; break; // 66
708 case X86II::XS: VEX_PP = 0x2; break; // F3
709 case X86II::XD: VEX_PP = 0x3; break; // F2
713 uint8_t EVEX_U = 1; // Always '1' so far
716 uint8_t EVEX_z = (HasEVEX_K && (TSFlags & X86II::EVEX_Z)) ? 1 : 0;
719 uint8_t EVEX_b = (TSFlags & X86II::EVEX_B) ? 1 : 0;
725 uint8_t EVEX_aaa = 0;
727 bool EncodeRC = false;
729 // Classify VEX_B, VEX_4V, VEX_R, VEX_X
730 unsigned NumOps = Desc.getNumOperands();
731 unsigned CurOp = X86II::getOperandBias(Desc);
733 switch (TSFlags & X86II::FormMask) {
734 default: llvm_unreachable("Unexpected form in EmitVEXOpcodePrefix!");
737 case X86II::MRMDestMem: {
738 // MRMDestMem instructions forms:
739 // MemAddr, src1(ModR/M)
740 // MemAddr, src1(VEX_4V), src2(ModR/M)
741 // MemAddr, src1(ModR/M), imm8
743 unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
744 VEX_B = ~(BaseRegEnc >> 3) & 1;
745 unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
746 VEX_X = ~(IndexRegEnc >> 3) & 1;
747 if (!HasVEX_4V) // Only needed with VSIB which don't use VVVV.
748 EVEX_V2 = ~(IndexRegEnc >> 4) & 1;
750 CurOp += X86::AddrNumOperands;
753 EVEX_aaa = getX86RegEncoding(MI, CurOp++);
756 unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
757 VEX_4V = ~VRegEnc & 0xf;
758 EVEX_V2 = ~(VRegEnc >> 4) & 1;
761 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
762 VEX_R = ~(RegEnc >> 3) & 1;
763 EVEX_R2 = ~(RegEnc >> 4) & 1;
766 case X86II::MRMSrcMem: {
767 // MRMSrcMem instructions forms:
768 // src1(ModR/M), MemAddr
769 // src1(ModR/M), src2(VEX_4V), MemAddr
770 // src1(ModR/M), MemAddr, imm8
771 // src1(ModR/M), MemAddr, src2(Imm[7:4])
774 // dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(Imm[7:4])
775 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
776 VEX_R = ~(RegEnc >> 3) & 1;
777 EVEX_R2 = ~(RegEnc >> 4) & 1;
780 EVEX_aaa = getX86RegEncoding(MI, CurOp++);
783 unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
784 VEX_4V = ~VRegEnc & 0xf;
785 EVEX_V2 = ~(VRegEnc >> 4) & 1;
788 unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
789 VEX_B = ~(BaseRegEnc >> 3) & 1;
790 unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
791 VEX_X = ~(IndexRegEnc >> 3) & 1;
792 if (!HasVEX_4V) // Only needed with VSIB which don't use VVVV.
793 EVEX_V2 = ~(IndexRegEnc >> 4) & 1;
797 case X86II::MRMSrcMem4VOp3: {
798 // Instruction format for 4VOp3:
799 // src1(ModR/M), MemAddr, src3(VEX_4V)
800 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
801 VEX_R = ~(RegEnc >> 3) & 1;
803 unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
804 VEX_B = ~(BaseRegEnc >> 3) & 1;
805 unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
806 VEX_X = ~(IndexRegEnc >> 3) & 1;
808 VEX_4V = ~getX86RegEncoding(MI, CurOp + X86::AddrNumOperands) & 0xf;
811 case X86II::MRMSrcMemOp4: {
812 // dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
813 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
814 VEX_R = ~(RegEnc >> 3) & 1;
816 unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
817 VEX_4V = ~VRegEnc & 0xf;
819 unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
820 VEX_B = ~(BaseRegEnc >> 3) & 1;
821 unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
822 VEX_X = ~(IndexRegEnc >> 3) & 1;
825 case X86II::MRM0m: case X86II::MRM1m:
826 case X86II::MRM2m: case X86II::MRM3m:
827 case X86II::MRM4m: case X86II::MRM5m:
828 case X86II::MRM6m: case X86II::MRM7m: {
829 // MRM[0-9]m instructions forms:
831 // src1(VEX_4V), MemAddr
833 unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
834 VEX_4V = ~VRegEnc & 0xf;
835 EVEX_V2 = ~(VRegEnc >> 4) & 1;
839 EVEX_aaa = getX86RegEncoding(MI, CurOp++);
841 unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
842 VEX_B = ~(BaseRegEnc >> 3) & 1;
843 unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
844 VEX_X = ~(IndexRegEnc >> 3) & 1;
847 case X86II::MRMSrcReg: {
848 // MRMSrcReg instructions forms:
849 // dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(Imm[7:4])
850 // dst(ModR/M), src1(ModR/M)
851 // dst(ModR/M), src1(ModR/M), imm8
854 // dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
855 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
856 VEX_R = ~(RegEnc >> 3) & 1;
857 EVEX_R2 = ~(RegEnc >> 4) & 1;
860 EVEX_aaa = getX86RegEncoding(MI, CurOp++);
863 unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
864 VEX_4V = ~VRegEnc & 0xf;
865 EVEX_V2 = ~(VRegEnc >> 4) & 1;
868 RegEnc = getX86RegEncoding(MI, CurOp++);
869 VEX_B = ~(RegEnc >> 3) & 1;
870 VEX_X = ~(RegEnc >> 4) & 1;
874 unsigned RcOperand = NumOps-1;
875 assert(RcOperand >= CurOp);
876 EVEX_rc = MI.getOperand(RcOperand).getImm() & 0x3;
882 case X86II::MRMSrcReg4VOp3: {
883 // Instruction format for 4VOp3:
884 // src1(ModR/M), src2(ModR/M), src3(VEX_4V)
885 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
886 VEX_R = ~(RegEnc >> 3) & 1;
888 RegEnc = getX86RegEncoding(MI, CurOp++);
889 VEX_B = ~(RegEnc >> 3) & 1;
891 VEX_4V = ~getX86RegEncoding(MI, CurOp++) & 0xf;
894 case X86II::MRMSrcRegOp4: {
895 // dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
896 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
897 VEX_R = ~(RegEnc >> 3) & 1;
899 unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
900 VEX_4V = ~VRegEnc & 0xf;
902 // Skip second register source (encoded in Imm[7:4])
905 RegEnc = getX86RegEncoding(MI, CurOp++);
906 VEX_B = ~(RegEnc >> 3) & 1;
907 VEX_X = ~(RegEnc >> 4) & 1;
910 case X86II::MRMDestReg: {
911 // MRMDestReg instructions forms:
912 // dst(ModR/M), src(ModR/M)
913 // dst(ModR/M), src(ModR/M), imm8
914 // dst(ModR/M), src1(VEX_4V), src2(ModR/M)
915 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
916 VEX_B = ~(RegEnc >> 3) & 1;
917 VEX_X = ~(RegEnc >> 4) & 1;
920 EVEX_aaa = getX86RegEncoding(MI, CurOp++);
923 unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
924 VEX_4V = ~VRegEnc & 0xf;
925 EVEX_V2 = ~(VRegEnc >> 4) & 1;
928 RegEnc = getX86RegEncoding(MI, CurOp++);
929 VEX_R = ~(RegEnc >> 3) & 1;
930 EVEX_R2 = ~(RegEnc >> 4) & 1;
935 case X86II::MRM0r: case X86II::MRM1r:
936 case X86II::MRM2r: case X86II::MRM3r:
937 case X86II::MRM4r: case X86II::MRM5r:
938 case X86II::MRM6r: case X86II::MRM7r: {
939 // MRM0r-MRM7r instructions forms:
940 // dst(VEX_4V), src(ModR/M), imm8
942 unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
943 VEX_4V = ~VRegEnc & 0xf;
944 EVEX_V2 = ~(VRegEnc >> 4) & 1;
947 EVEX_aaa = getX86RegEncoding(MI, CurOp++);
949 unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
950 VEX_B = ~(RegEnc >> 3) & 1;
951 VEX_X = ~(RegEnc >> 4) & 1;
956 if (Encoding == X86II::VEX || Encoding == X86II::XOP) {
957 // VEX opcode prefix can have 2 or 3 bytes
960 // +-----+ +--------------+ +-------------------+
961 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
962 // +-----+ +--------------+ +-------------------+
964 // +-----+ +-------------------+
965 // | C5h | | R | vvvv | L | pp |
966 // +-----+ +-------------------+
968 // XOP uses a similar prefix:
969 // +-----+ +--------------+ +-------------------+
970 // | 8Fh | | RXB | m-mmmm | | W | vvvv | L | pp |
971 // +-----+ +--------------+ +-------------------+
972 uint8_t LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
974 // Can we use the 2 byte VEX prefix?
975 if (Encoding == X86II::VEX && VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) {
976 EmitByte(0xC5, CurByte, OS);
977 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
982 EmitByte(Encoding == X86II::XOP ? 0x8F : 0xC4, CurByte, OS);
983 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
984 EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
986 assert(Encoding == X86II::EVEX && "unknown encoding!");
987 // EVEX opcode prefix can have 4 bytes
989 // +-----+ +--------------+ +-------------------+ +------------------------+
990 // | 62h | | RXBR' | 00mm | | W | vvvv | U | pp | | z | L'L | b | v' | aaa |
991 // +-----+ +--------------+ +-------------------+ +------------------------+
992 assert((VEX_5M & 0x3) == VEX_5M
993 && "More than 2 significant bits in VEX.m-mmmm fields for EVEX!");
995 EmitByte(0x62, CurByte, OS);
996 EmitByte((VEX_R << 7) |
1000 VEX_5M, CurByte, OS);
1001 EmitByte((VEX_W << 7) |
1004 VEX_PP, CurByte, OS);
1006 EmitByte((EVEX_z << 7) |
1010 EVEX_aaa, CurByte, OS);
1012 EmitByte((EVEX_z << 7) |
1017 EVEX_aaa, CurByte, OS);
1021 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
1022 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
1023 /// size, and 3) use of X86-64 extended registers.
1024 uint8_t X86MCCodeEmitter::DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
1026 const MCInstrDesc &Desc) const {
1028 bool UsesHighByteReg = false;
1030 if (TSFlags & X86II::REX_W)
1031 REX |= 1 << 3; // set REX.W
1033 if (MI.getNumOperands() == 0) return REX;
1035 unsigned NumOps = MI.getNumOperands();
1036 unsigned CurOp = X86II::getOperandBias(Desc);
1038 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
1039 for (unsigned i = CurOp; i != NumOps; ++i) {
1040 const MCOperand &MO = MI.getOperand(i);
1041 if (!MO.isReg()) continue;
1042 unsigned Reg = MO.getReg();
1043 if (Reg == X86::AH || Reg == X86::BH || Reg == X86::CH || Reg == X86::DH)
1044 UsesHighByteReg = true;
1045 if (X86II::isX86_64NonExtLowByteReg(Reg))
1046 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
1047 // that returns non-zero.
1048 REX |= 0x40; // REX fixed encoding prefix
1051 switch (TSFlags & X86II::FormMask) {
1052 case X86II::AddRegFrm:
1053 REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1055 case X86II::MRMSrcReg:
1056 REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1057 REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1059 case X86II::MRMSrcMem: {
1060 REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1061 REX |= isREXExtendedReg(MI, MemOperand+X86::AddrBaseReg) << 0; // REX.B
1062 REX |= isREXExtendedReg(MI, MemOperand+X86::AddrIndexReg) << 1; // REX.X
1063 CurOp += X86::AddrNumOperands;
1066 case X86II::MRMDestReg:
1067 REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1068 REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1070 case X86II::MRMDestMem:
1071 REX |= isREXExtendedReg(MI, MemOperand+X86::AddrBaseReg) << 0; // REX.B
1072 REX |= isREXExtendedReg(MI, MemOperand+X86::AddrIndexReg) << 1; // REX.X
1073 CurOp += X86::AddrNumOperands;
1074 REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1077 case X86II::MRM0m: case X86II::MRM1m:
1078 case X86II::MRM2m: case X86II::MRM3m:
1079 case X86II::MRM4m: case X86II::MRM5m:
1080 case X86II::MRM6m: case X86II::MRM7m:
1081 REX |= isREXExtendedReg(MI, MemOperand+X86::AddrBaseReg) << 0; // REX.B
1082 REX |= isREXExtendedReg(MI, MemOperand+X86::AddrIndexReg) << 1; // REX.X
1085 case X86II::MRM0r: case X86II::MRM1r:
1086 case X86II::MRM2r: case X86II::MRM3r:
1087 case X86II::MRM4r: case X86II::MRM5r:
1088 case X86II::MRM6r: case X86II::MRM7r:
1089 REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1092 if (REX && UsesHighByteReg)
1093 report_fatal_error("Cannot encode high byte register in REX-prefixed instruction");
1098 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
1099 void X86MCCodeEmitter::EmitSegmentOverridePrefix(unsigned &CurByte,
1100 unsigned SegOperand,
1102 raw_ostream &OS) const {
1103 // Check for explicit segment override on memory operand.
1104 switch (MI.getOperand(SegOperand).getReg()) {
1105 default: llvm_unreachable("Unknown segment register!");
1107 case X86::CS: EmitByte(0x2E, CurByte, OS); break;
1108 case X86::SS: EmitByte(0x36, CurByte, OS); break;
1109 case X86::DS: EmitByte(0x3E, CurByte, OS); break;
1110 case X86::ES: EmitByte(0x26, CurByte, OS); break;
1111 case X86::FS: EmitByte(0x64, CurByte, OS); break;
1112 case X86::GS: EmitByte(0x65, CurByte, OS); break;
1116 /// Emit all instruction prefixes prior to the opcode.
1118 /// MemOperand is the operand # of the start of a memory operand if present. If
1119 /// Not present, it is -1.
1121 /// Returns true if a REX prefix was used.
1122 bool X86MCCodeEmitter::emitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
1123 int MemOperand, const MCInst &MI,
1124 const MCInstrDesc &Desc,
1125 const MCSubtargetInfo &STI,
1126 raw_ostream &OS) const {
1128 // Emit the operand size opcode prefix as needed.
1129 if ((TSFlags & X86II::OpSizeMask) == (is16BitMode(STI) ? X86II::OpSize32
1131 EmitByte(0x66, CurByte, OS);
1133 // Emit the LOCK opcode prefix.
1134 if (TSFlags & X86II::LOCK || MI.getFlags() & X86::IP_HAS_LOCK)
1135 EmitByte(0xF0, CurByte, OS);
1137 switch (TSFlags & X86II::OpPrefixMask) {
1138 case X86II::PD: // 66
1139 EmitByte(0x66, CurByte, OS);
1141 case X86II::XS: // F3
1142 EmitByte(0xF3, CurByte, OS);
1144 case X86II::XD: // F2
1145 EmitByte(0xF2, CurByte, OS);
1149 // Handle REX prefix.
1150 // FIXME: Can this come before F2 etc to simplify emission?
1151 if (is64BitMode(STI)) {
1152 if (uint8_t REX = DetermineREXPrefix(MI, TSFlags, MemOperand, Desc)) {
1153 EmitByte(0x40 | REX, CurByte, OS);
1157 assert(!(TSFlags & X86II::REX_W) && "REX.W requires 64bit mode.");
1160 // 0x0F escape code must be emitted just before the opcode.
1161 switch (TSFlags & X86II::OpMapMask) {
1162 case X86II::TB: // Two-byte opcode map
1163 case X86II::T8: // 0F 38
1164 case X86II::TA: // 0F 3A
1165 EmitByte(0x0F, CurByte, OS);
1169 switch (TSFlags & X86II::OpMapMask) {
1170 case X86II::T8: // 0F 38
1171 EmitByte(0x38, CurByte, OS);
1173 case X86II::TA: // 0F 3A
1174 EmitByte(0x3A, CurByte, OS);
1180 void X86MCCodeEmitter::
1181 encodeInstruction(const MCInst &MI, raw_ostream &OS,
1182 SmallVectorImpl<MCFixup> &Fixups,
1183 const MCSubtargetInfo &STI) const {
1184 unsigned Opcode = MI.getOpcode();
1185 const MCInstrDesc &Desc = MCII.get(Opcode);
1186 uint64_t TSFlags = Desc.TSFlags;
1187 unsigned Flags = MI.getFlags();
1189 // Pseudo instructions don't get encoded.
1190 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
1193 unsigned NumOps = Desc.getNumOperands();
1194 unsigned CurOp = X86II::getOperandBias(Desc);
1196 // Keep track of the current byte being emitted.
1197 unsigned CurByte = 0;
1199 // Encoding type for this instruction.
1200 uint64_t Encoding = TSFlags & X86II::EncodingMask;
1202 // It uses the VEX.VVVV field?
1203 bool HasVEX_4V = TSFlags & X86II::VEX_4V;
1204 bool HasVEX_I8Reg = (TSFlags & X86II::ImmMask) == X86II::Imm8Reg;
1206 // It uses the EVEX.aaa field?
1207 bool HasEVEX_K = TSFlags & X86II::EVEX_K;
1208 bool HasEVEX_RC = TSFlags & X86II::EVEX_RC;
1210 // Used if a register is encoded in 7:4 of immediate.
1211 unsigned I8RegNum = 0;
1213 // Determine where the memory operand starts, if present.
1214 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags);
1215 if (MemoryOperand != -1) MemoryOperand += CurOp;
1217 // Emit segment override opcode prefix as needed.
1218 if (MemoryOperand >= 0)
1219 EmitSegmentOverridePrefix(CurByte, MemoryOperand+X86::AddrSegmentReg,
1222 // Emit the repeat opcode prefix as needed.
1223 if (TSFlags & X86II::REP || Flags & X86::IP_HAS_REPEAT)
1224 EmitByte(0xF3, CurByte, OS);
1225 if (Flags & X86::IP_HAS_REPEAT_NE)
1226 EmitByte(0xF2, CurByte, OS);
1228 // Emit the address size opcode prefix as needed.
1229 bool need_address_override;
1230 uint64_t AdSize = TSFlags & X86II::AdSizeMask;
1231 if ((is16BitMode(STI) && AdSize == X86II::AdSize32) ||
1232 (is32BitMode(STI) && AdSize == X86II::AdSize16) ||
1233 (is64BitMode(STI) && AdSize == X86II::AdSize32)) {
1234 need_address_override = true;
1235 } else if (MemoryOperand < 0) {
1236 need_address_override = false;
1237 } else if (is64BitMode(STI)) {
1238 assert(!Is16BitMemOperand(MI, MemoryOperand, STI));
1239 need_address_override = Is32BitMemOperand(MI, MemoryOperand);
1240 } else if (is32BitMode(STI)) {
1241 assert(!Is64BitMemOperand(MI, MemoryOperand));
1242 need_address_override = Is16BitMemOperand(MI, MemoryOperand, STI);
1244 assert(is16BitMode(STI));
1245 assert(!Is64BitMemOperand(MI, MemoryOperand));
1246 need_address_override = !Is16BitMemOperand(MI, MemoryOperand, STI);
1249 if (need_address_override)
1250 EmitByte(0x67, CurByte, OS);
1254 Rex = emitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, STI, OS);
1256 EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
1258 uint8_t BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
1260 if (TSFlags & X86II::Has3DNow0F0FOpcode)
1261 BaseOpcode = 0x0F; // Weird 3DNow! encoding.
1263 uint64_t Form = TSFlags & X86II::FormMask;
1265 default: errs() << "FORM: " << Form << "\n";
1266 llvm_unreachable("Unknown FormMask value in X86MCCodeEmitter!");
1268 llvm_unreachable("Pseudo instruction shouldn't be emitted");
1269 case X86II::RawFrmDstSrc: {
1270 unsigned siReg = MI.getOperand(1).getReg();
1271 assert(((siReg == X86::SI && MI.getOperand(0).getReg() == X86::DI) ||
1272 (siReg == X86::ESI && MI.getOperand(0).getReg() == X86::EDI) ||
1273 (siReg == X86::RSI && MI.getOperand(0).getReg() == X86::RDI)) &&
1274 "SI and DI register sizes do not match");
1275 // Emit segment override opcode prefix as needed (not for %ds).
1276 if (MI.getOperand(2).getReg() != X86::DS)
1277 EmitSegmentOverridePrefix(CurByte, 2, MI, OS);
1278 // Emit AdSize prefix as needed.
1279 if ((!is32BitMode(STI) && siReg == X86::ESI) ||
1280 (is32BitMode(STI) && siReg == X86::SI))
1281 EmitByte(0x67, CurByte, OS);
1282 CurOp += 3; // Consume operands.
1283 EmitByte(BaseOpcode, CurByte, OS);
1286 case X86II::RawFrmSrc: {
1287 unsigned siReg = MI.getOperand(0).getReg();
1288 // Emit segment override opcode prefix as needed (not for %ds).
1289 if (MI.getOperand(1).getReg() != X86::DS)
1290 EmitSegmentOverridePrefix(CurByte, 1, 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 += 2; // Consume operands.
1296 EmitByte(BaseOpcode, CurByte, OS);
1299 case X86II::RawFrmDst: {
1300 unsigned siReg = MI.getOperand(0).getReg();
1301 // Emit AdSize prefix as needed.
1302 if ((!is32BitMode(STI) && siReg == X86::EDI) ||
1303 (is32BitMode(STI) && siReg == X86::DI))
1304 EmitByte(0x67, CurByte, OS);
1305 ++CurOp; // Consume operand.
1306 EmitByte(BaseOpcode, CurByte, OS);
1309 case X86II::RawFrm: {
1310 EmitByte(BaseOpcode, CurByte, OS);
1312 if (!is64BitMode(STI) || !isPCRel32Branch(MI))
1315 const MCOperand &Op = MI.getOperand(CurOp++);
1316 EmitImmediate(Op, MI.getLoc(), X86II::getSizeOfImm(TSFlags),
1317 MCFixupKind(X86::reloc_branch_4byte_pcrel), CurByte, OS,
1321 case X86II::RawFrmMemOffs:
1322 // Emit segment override opcode prefix as needed.
1323 EmitSegmentOverridePrefix(CurByte, 1, MI, OS);
1324 EmitByte(BaseOpcode, CurByte, OS);
1325 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1326 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1327 CurByte, OS, Fixups);
1328 ++CurOp; // skip segment operand
1330 case X86II::RawFrmImm8:
1331 EmitByte(BaseOpcode, CurByte, OS);
1332 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1333 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1334 CurByte, OS, Fixups);
1335 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 1, FK_Data_1, CurByte,
1338 case X86II::RawFrmImm16:
1339 EmitByte(BaseOpcode, CurByte, OS);
1340 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1341 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1342 CurByte, OS, Fixups);
1343 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 2, FK_Data_2, CurByte,
1347 case X86II::AddRegFrm:
1348 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
1351 case X86II::MRMDestReg: {
1352 EmitByte(BaseOpcode, CurByte, OS);
1353 unsigned SrcRegNum = CurOp + 1;
1355 if (HasEVEX_K) // Skip writemask
1358 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1361 EmitRegModRMByte(MI.getOperand(CurOp),
1362 GetX86RegNum(MI.getOperand(SrcRegNum)), CurByte, OS);
1363 CurOp = SrcRegNum + 1;
1366 case X86II::MRMDestMem: {
1367 EmitByte(BaseOpcode, CurByte, OS);
1368 unsigned SrcRegNum = CurOp + X86::AddrNumOperands;
1370 if (HasEVEX_K) // Skip writemask
1373 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1376 emitMemModRMByte(MI, CurOp, GetX86RegNum(MI.getOperand(SrcRegNum)), TSFlags,
1377 Rex, CurByte, OS, Fixups, STI);
1378 CurOp = SrcRegNum + 1;
1381 case X86II::MRMSrcReg: {
1382 EmitByte(BaseOpcode, CurByte, OS);
1383 unsigned SrcRegNum = CurOp + 1;
1385 if (HasEVEX_K) // Skip writemask
1388 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1391 EmitRegModRMByte(MI.getOperand(SrcRegNum),
1392 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1393 CurOp = SrcRegNum + 1;
1395 I8RegNum = getX86RegEncoding(MI, CurOp++);
1396 // do not count the rounding control operand
1401 case X86II::MRMSrcReg4VOp3: {
1402 EmitByte(BaseOpcode, CurByte, OS);
1403 unsigned SrcRegNum = CurOp + 1;
1405 EmitRegModRMByte(MI.getOperand(SrcRegNum),
1406 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1407 CurOp = SrcRegNum + 1;
1408 ++CurOp; // Encoded in VEX.VVVV
1411 case X86II::MRMSrcRegOp4: {
1412 EmitByte(BaseOpcode, CurByte, OS);
1413 unsigned SrcRegNum = CurOp + 1;
1415 // Skip 1st src (which is encoded in VEX_VVVV)
1418 // Capture 2nd src (which is encoded in Imm[7:4])
1419 assert(HasVEX_I8Reg && "MRMSrcRegOp4 should imply VEX_I8Reg");
1420 I8RegNum = getX86RegEncoding(MI, SrcRegNum++);
1422 EmitRegModRMByte(MI.getOperand(SrcRegNum),
1423 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1424 CurOp = SrcRegNum + 1;
1427 case X86II::MRMSrcMem: {
1428 unsigned FirstMemOp = CurOp+1;
1430 if (HasEVEX_K) // Skip writemask
1434 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
1436 EmitByte(BaseOpcode, CurByte, OS);
1438 emitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1439 TSFlags, Rex, CurByte, OS, Fixups, STI);
1440 CurOp = FirstMemOp + X86::AddrNumOperands;
1442 I8RegNum = getX86RegEncoding(MI, CurOp++);
1445 case X86II::MRMSrcMem4VOp3: {
1446 unsigned FirstMemOp = CurOp+1;
1448 EmitByte(BaseOpcode, CurByte, OS);
1450 emitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1451 TSFlags, Rex, CurByte, OS, Fixups, STI);
1452 CurOp = FirstMemOp + X86::AddrNumOperands;
1453 ++CurOp; // Encoded in VEX.VVVV.
1456 case X86II::MRMSrcMemOp4: {
1457 unsigned FirstMemOp = CurOp+1;
1459 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
1461 // Capture second register source (encoded in Imm[7:4])
1462 assert(HasVEX_I8Reg && "MRMSrcRegOp4 should imply VEX_I8Reg");
1463 I8RegNum = getX86RegEncoding(MI, FirstMemOp++);
1465 EmitByte(BaseOpcode, CurByte, OS);
1467 emitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1468 TSFlags, Rex, CurByte, OS, Fixups, STI);
1469 CurOp = FirstMemOp + X86::AddrNumOperands;
1474 case X86II::MRM0r: case X86II::MRM1r:
1475 case X86II::MRM2r: case X86II::MRM3r:
1476 case X86II::MRM4r: case X86II::MRM5r:
1477 case X86II::MRM6r: case X86II::MRM7r:
1478 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1480 if (HasEVEX_K) // Skip writemask
1482 EmitByte(BaseOpcode, CurByte, OS);
1483 EmitRegModRMByte(MI.getOperand(CurOp++),
1484 (Form == X86II::MRMXr) ? 0 : Form-X86II::MRM0r,
1489 case X86II::MRM0m: case X86II::MRM1m:
1490 case X86II::MRM2m: case X86II::MRM3m:
1491 case X86II::MRM4m: case X86II::MRM5m:
1492 case X86II::MRM6m: case X86II::MRM7m:
1493 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1495 if (HasEVEX_K) // Skip writemask
1497 EmitByte(BaseOpcode, CurByte, OS);
1498 emitMemModRMByte(MI, CurOp,
1499 (Form == X86II::MRMXm) ? 0 : Form - X86II::MRM0m, TSFlags,
1500 Rex, CurByte, OS, Fixups, STI);
1501 CurOp += X86::AddrNumOperands;
1504 case X86II::MRM_C0: case X86II::MRM_C1: case X86II::MRM_C2:
1505 case X86II::MRM_C3: case X86II::MRM_C4: case X86II::MRM_C5:
1506 case X86II::MRM_C6: case X86II::MRM_C7: case X86II::MRM_C8:
1507 case X86II::MRM_C9: case X86II::MRM_CA: case X86II::MRM_CB:
1508 case X86II::MRM_CC: case X86II::MRM_CD: case X86II::MRM_CE:
1509 case X86II::MRM_CF: case X86II::MRM_D0: case X86II::MRM_D1:
1510 case X86II::MRM_D2: case X86II::MRM_D3: case X86II::MRM_D4:
1511 case X86II::MRM_D5: case X86II::MRM_D6: case X86II::MRM_D7:
1512 case X86II::MRM_D8: case X86II::MRM_D9: case X86II::MRM_DA:
1513 case X86II::MRM_DB: case X86II::MRM_DC: case X86II::MRM_DD:
1514 case X86II::MRM_DE: case X86II::MRM_DF: case X86II::MRM_E0:
1515 case X86II::MRM_E1: case X86II::MRM_E2: case X86II::MRM_E3:
1516 case X86II::MRM_E4: case X86II::MRM_E5: case X86II::MRM_E6:
1517 case X86II::MRM_E7: case X86II::MRM_E8: case X86II::MRM_E9:
1518 case X86II::MRM_EA: case X86II::MRM_EB: case X86II::MRM_EC:
1519 case X86II::MRM_ED: case X86II::MRM_EE: case X86II::MRM_EF:
1520 case X86II::MRM_F0: case X86II::MRM_F1: case X86II::MRM_F2:
1521 case X86II::MRM_F3: case X86II::MRM_F4: case X86II::MRM_F5:
1522 case X86II::MRM_F6: case X86II::MRM_F7: case X86II::MRM_F8:
1523 case X86II::MRM_F9: case X86II::MRM_FA: case X86II::MRM_FB:
1524 case X86II::MRM_FC: case X86II::MRM_FD: case X86II::MRM_FE:
1526 EmitByte(BaseOpcode, CurByte, OS);
1527 EmitByte(0xC0 + Form - X86II::MRM_C0, CurByte, OS);
1532 // The last source register of a 4 operand instruction in AVX is encoded
1533 // in bits[7:4] of a immediate byte.
1534 assert(I8RegNum < 16 && "Register encoding out of range");
1536 if (CurOp != NumOps) {
1537 unsigned Val = MI.getOperand(CurOp++).getImm();
1538 assert(Val < 16 && "Immediate operand value out of range");
1541 EmitImmediate(MCOperand::createImm(I8RegNum), MI.getLoc(), 1, FK_Data_1,
1542 CurByte, OS, Fixups);
1544 // If there is a remaining operand, it must be a trailing immediate. Emit it
1545 // according to the right size for the instruction. Some instructions
1546 // (SSE4a extrq and insertq) have two trailing immediates.
1547 while (CurOp != NumOps && NumOps - CurOp <= 2) {
1548 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1549 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1550 CurByte, OS, Fixups);
1554 if (TSFlags & X86II::Has3DNow0F0FOpcode)
1555 EmitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS);
1559 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
1560 errs() << "Cannot encode all operands of: ";
1568 MCCodeEmitter *llvm::createX86MCCodeEmitter(const MCInstrInfo &MCII,
1569 const MCRegisterInfo &MRI,
1571 return new X86MCCodeEmitter(MCII, Ctx);