xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp (revision 8bcb0991864975618c09697b1aca10683346d9f0)
1 //===-- X86MCCodeEmitter.cpp - Convert X86 code to machine code -----------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the X86MCCodeEmitter class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "MCTargetDesc/X86BaseInfo.h"
14 #include "MCTargetDesc/X86FixupKinds.h"
15 #include "MCTargetDesc/X86MCTargetDesc.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/MC/MCCodeEmitter.h"
18 #include "llvm/MC/MCContext.h"
19 #include "llvm/MC/MCExpr.h"
20 #include "llvm/MC/MCFixup.h"
21 #include "llvm/MC/MCInst.h"
22 #include "llvm/MC/MCInstrDesc.h"
23 #include "llvm/MC/MCInstrInfo.h"
24 #include "llvm/MC/MCRegisterInfo.h"
25 #include "llvm/MC/MCSubtargetInfo.h"
26 #include "llvm/MC/MCSymbol.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include <cassert>
30 #include <cstdint>
31 #include <cstdlib>
32 
33 using namespace llvm;
34 
35 #define DEBUG_TYPE "mccodeemitter"
36 
37 namespace {
38 
39 class X86MCCodeEmitter : public MCCodeEmitter {
40   const MCInstrInfo &MCII;
41   MCContext &Ctx;
42 
43 public:
44   X86MCCodeEmitter(const MCInstrInfo &mcii, MCContext &ctx)
45     : MCII(mcii), Ctx(ctx) {
46   }
47   X86MCCodeEmitter(const X86MCCodeEmitter &) = delete;
48   X86MCCodeEmitter &operator=(const X86MCCodeEmitter &) = delete;
49   ~X86MCCodeEmitter() override = default;
50 
51   bool is64BitMode(const MCSubtargetInfo &STI) const {
52     return STI.getFeatureBits()[X86::Mode64Bit];
53   }
54 
55   bool is32BitMode(const MCSubtargetInfo &STI) const {
56     return STI.getFeatureBits()[X86::Mode32Bit];
57   }
58 
59   bool is16BitMode(const MCSubtargetInfo &STI) const {
60     return STI.getFeatureBits()[X86::Mode16Bit];
61   }
62 
63   /// Is16BitMemOperand - Return true if the specified instruction has
64   /// a 16-bit memory operand. Op specifies the operand # of the memoperand.
65   bool Is16BitMemOperand(const MCInst &MI, unsigned Op,
66                          const MCSubtargetInfo &STI) const {
67     const MCOperand &BaseReg  = MI.getOperand(Op+X86::AddrBaseReg);
68     const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
69     const MCOperand &Disp     = MI.getOperand(Op+X86::AddrDisp);
70 
71     if (is16BitMode(STI) && BaseReg.getReg() == 0 &&
72         Disp.isImm() && Disp.getImm() < 0x10000)
73       return true;
74     if ((BaseReg.getReg() != 0 &&
75          X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) ||
76         (IndexReg.getReg() != 0 &&
77          X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg())))
78       return true;
79     return false;
80   }
81 
82   unsigned GetX86RegNum(const MCOperand &MO) const {
83     return Ctx.getRegisterInfo()->getEncodingValue(MO.getReg()) & 0x7;
84   }
85 
86   unsigned getX86RegEncoding(const MCInst &MI, unsigned OpNum) const {
87     return Ctx.getRegisterInfo()->getEncodingValue(
88                                                  MI.getOperand(OpNum).getReg());
89   }
90 
91   // Does this register require a bit to be set in REX prefix.
92   bool isREXExtendedReg(const MCInst &MI, unsigned OpNum) const {
93     return (getX86RegEncoding(MI, OpNum) >> 3) & 1;
94   }
95 
96   void EmitByte(uint8_t C, unsigned &CurByte, raw_ostream &OS) const {
97     OS << (char)C;
98     ++CurByte;
99   }
100 
101   void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
102                     raw_ostream &OS) const {
103     // Output the constant in little endian byte order.
104     for (unsigned i = 0; i != Size; ++i) {
105       EmitByte(Val & 255, CurByte, OS);
106       Val >>= 8;
107     }
108   }
109 
110   void EmitImmediate(const MCOperand &Disp, SMLoc Loc,
111                      unsigned ImmSize, MCFixupKind FixupKind,
112                      unsigned &CurByte, raw_ostream &OS,
113                      SmallVectorImpl<MCFixup> &Fixups,
114                      int ImmOffset = 0) const;
115 
116   static uint8_t ModRMByte(unsigned Mod, unsigned RegOpcode, unsigned RM) {
117     assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
118     return RM | (RegOpcode << 3) | (Mod << 6);
119   }
120 
121   void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
122                         unsigned &CurByte, raw_ostream &OS) const {
123     EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS);
124   }
125 
126   void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base,
127                    unsigned &CurByte, raw_ostream &OS) const {
128     // SIB byte is in the same format as the ModRMByte.
129     EmitByte(ModRMByte(SS, Index, Base), CurByte, OS);
130   }
131 
132   void emitMemModRMByte(const MCInst &MI, unsigned Op, unsigned RegOpcodeField,
133                         uint64_t TSFlags, bool Rex, unsigned &CurByte,
134                         raw_ostream &OS, SmallVectorImpl<MCFixup> &Fixups,
135                         const MCSubtargetInfo &STI) const;
136 
137   void encodeInstruction(const MCInst &MI, raw_ostream &OS,
138                          SmallVectorImpl<MCFixup> &Fixups,
139                          const MCSubtargetInfo &STI) const override;
140 
141   void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
142                            const MCInst &MI, const MCInstrDesc &Desc,
143                            raw_ostream &OS) const;
144 
145   void EmitSegmentOverridePrefix(unsigned &CurByte, unsigned SegOperand,
146                                  const MCInst &MI, raw_ostream &OS) const;
147 
148   bool emitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
149                         const MCInst &MI, const MCInstrDesc &Desc,
150                         const MCSubtargetInfo &STI, raw_ostream &OS) const;
151 
152   uint8_t DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
153                              int MemOperand, const MCInstrDesc &Desc) const;
154 
155   bool isPCRel32Branch(const MCInst &MI) const;
156 };
157 
158 } // end anonymous namespace
159 
160 /// isDisp8 - Return true if this signed displacement fits in a 8-bit
161 /// sign-extended field.
162 static bool isDisp8(int Value) {
163   return Value == (int8_t)Value;
164 }
165 
166 /// isCDisp8 - Return true if this signed displacement fits in a 8-bit
167 /// compressed dispacement field.
168 static bool isCDisp8(uint64_t TSFlags, int Value, int& CValue) {
169   assert(((TSFlags & X86II::EncodingMask) == X86II::EVEX) &&
170          "Compressed 8-bit displacement is only valid for EVEX inst.");
171 
172   unsigned CD8_Scale =
173     (TSFlags & X86II::CD8_Scale_Mask) >> X86II::CD8_Scale_Shift;
174   if (CD8_Scale == 0) {
175     CValue = Value;
176     return isDisp8(Value);
177   }
178 
179   unsigned Mask = CD8_Scale - 1;
180   assert((CD8_Scale & Mask) == 0 && "Invalid memory object size.");
181   if (Value & Mask) // Unaligned offset
182     return false;
183   Value /= (int)CD8_Scale;
184   bool Ret = (Value == (int8_t)Value);
185 
186   if (Ret)
187     CValue = Value;
188   return Ret;
189 }
190 
191 /// getImmFixupKind - Return the appropriate fixup kind to use for an immediate
192 /// in an instruction with the specified TSFlags.
193 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
194   unsigned Size = X86II::getSizeOfImm(TSFlags);
195   bool isPCRel = X86II::isImmPCRel(TSFlags);
196 
197   if (X86II::isImmSigned(TSFlags)) {
198     switch (Size) {
199     default: llvm_unreachable("Unsupported signed fixup size!");
200     case 4: return MCFixupKind(X86::reloc_signed_4byte);
201     }
202   }
203   return MCFixup::getKindForSize(Size, isPCRel);
204 }
205 
206 /// Is32BitMemOperand - Return true if the specified instruction has
207 /// a 32-bit memory operand. Op specifies the operand # of the memoperand.
208 static bool Is32BitMemOperand(const MCInst &MI, unsigned Op) {
209   const MCOperand &BaseReg  = MI.getOperand(Op+X86::AddrBaseReg);
210   const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
211 
212   if ((BaseReg.getReg() != 0 &&
213        X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) ||
214       (IndexReg.getReg() != 0 &&
215        X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg())))
216     return true;
217   if (BaseReg.getReg() == X86::EIP) {
218     assert(IndexReg.getReg() == 0 && "Invalid eip-based address.");
219     return true;
220   }
221   if (IndexReg.getReg() == X86::EIZ)
222     return true;
223   return false;
224 }
225 
226 /// Is64BitMemOperand - Return true if the specified instruction has
227 /// a 64-bit memory operand. Op specifies the operand # of the memoperand.
228 #ifndef NDEBUG
229 static bool Is64BitMemOperand(const MCInst &MI, unsigned Op) {
230   const MCOperand &BaseReg  = MI.getOperand(Op+X86::AddrBaseReg);
231   const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
232 
233   if ((BaseReg.getReg() != 0 &&
234        X86MCRegisterClasses[X86::GR64RegClassID].contains(BaseReg.getReg())) ||
235       (IndexReg.getReg() != 0 &&
236        X86MCRegisterClasses[X86::GR64RegClassID].contains(IndexReg.getReg())))
237     return true;
238   return false;
239 }
240 #endif
241 
242 /// StartsWithGlobalOffsetTable - Check if this expression starts with
243 ///  _GLOBAL_OFFSET_TABLE_ and if it is of the form
244 ///  _GLOBAL_OFFSET_TABLE_-symbol. This is needed to support PIC on ELF
245 /// i386 as _GLOBAL_OFFSET_TABLE_ is magical. We check only simple case that
246 /// are know to be used: _GLOBAL_OFFSET_TABLE_ by itself or at the start
247 /// of a binary expression.
248 enum GlobalOffsetTableExprKind {
249   GOT_None,
250   GOT_Normal,
251   GOT_SymDiff
252 };
253 static GlobalOffsetTableExprKind
254 StartsWithGlobalOffsetTable(const MCExpr *Expr) {
255   const MCExpr *RHS = nullptr;
256   if (Expr->getKind() == MCExpr::Binary) {
257     const MCBinaryExpr *BE = static_cast<const MCBinaryExpr *>(Expr);
258     Expr = BE->getLHS();
259     RHS = BE->getRHS();
260   }
261 
262   if (Expr->getKind() != MCExpr::SymbolRef)
263     return GOT_None;
264 
265   const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
266   const MCSymbol &S = Ref->getSymbol();
267   if (S.getName() != "_GLOBAL_OFFSET_TABLE_")
268     return GOT_None;
269   if (RHS && RHS->getKind() == MCExpr::SymbolRef)
270     return GOT_SymDiff;
271   return GOT_Normal;
272 }
273 
274 static bool HasSecRelSymbolRef(const MCExpr *Expr) {
275   if (Expr->getKind() == MCExpr::SymbolRef) {
276     const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
277     return Ref->getKind() == MCSymbolRefExpr::VK_SECREL;
278   }
279   return false;
280 }
281 
282 bool X86MCCodeEmitter::isPCRel32Branch(const MCInst &MI) const {
283   unsigned Opcode = MI.getOpcode();
284   const MCInstrDesc &Desc = MCII.get(Opcode);
285   if ((Opcode != X86::CALL64pcrel32 && Opcode != X86::JMP_4) ||
286       getImmFixupKind(Desc.TSFlags) != FK_PCRel_4)
287     return false;
288 
289   unsigned CurOp = X86II::getOperandBias(Desc);
290   const MCOperand &Op = MI.getOperand(CurOp);
291   if (!Op.isExpr())
292     return false;
293 
294   const MCSymbolRefExpr *Ref = dyn_cast<MCSymbolRefExpr>(Op.getExpr());
295   return Ref && Ref->getKind() == MCSymbolRefExpr::VK_None;
296 }
297 
298 void X86MCCodeEmitter::
299 EmitImmediate(const MCOperand &DispOp, SMLoc Loc, unsigned Size,
300               MCFixupKind FixupKind, unsigned &CurByte, raw_ostream &OS,
301               SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
302   const MCExpr *Expr = nullptr;
303   if (DispOp.isImm()) {
304     // If this is a simple integer displacement that doesn't require a
305     // relocation, emit it now.
306     if (FixupKind != FK_PCRel_1 &&
307         FixupKind != FK_PCRel_2 &&
308         FixupKind != FK_PCRel_4) {
309       EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
310       return;
311     }
312     Expr = MCConstantExpr::create(DispOp.getImm(), Ctx);
313   } else {
314     Expr = DispOp.getExpr();
315   }
316 
317   // If we have an immoffset, add it to the expression.
318   if ((FixupKind == FK_Data_4 ||
319        FixupKind == FK_Data_8 ||
320        FixupKind == MCFixupKind(X86::reloc_signed_4byte))) {
321     GlobalOffsetTableExprKind Kind = StartsWithGlobalOffsetTable(Expr);
322     if (Kind != GOT_None) {
323       assert(ImmOffset == 0);
324 
325       if (Size == 8) {
326         FixupKind = MCFixupKind(X86::reloc_global_offset_table8);
327       } else {
328         assert(Size == 4);
329         FixupKind = MCFixupKind(X86::reloc_global_offset_table);
330       }
331 
332       if (Kind == GOT_Normal)
333         ImmOffset = CurByte;
334     } else if (Expr->getKind() == MCExpr::SymbolRef) {
335       if (HasSecRelSymbolRef(Expr)) {
336         FixupKind = MCFixupKind(FK_SecRel_4);
337       }
338     } else if (Expr->getKind() == MCExpr::Binary) {
339       const MCBinaryExpr *Bin = static_cast<const MCBinaryExpr*>(Expr);
340       if (HasSecRelSymbolRef(Bin->getLHS())
341           || HasSecRelSymbolRef(Bin->getRHS())) {
342         FixupKind = MCFixupKind(FK_SecRel_4);
343       }
344     }
345   }
346 
347   // If the fixup is pc-relative, we need to bias the value to be relative to
348   // the start of the field, not the end of the field.
349   if (FixupKind == FK_PCRel_4 ||
350       FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
351       FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load) ||
352       FixupKind == MCFixupKind(X86::reloc_riprel_4byte_relax) ||
353       FixupKind == MCFixupKind(X86::reloc_riprel_4byte_relax_rex) ||
354       FixupKind == MCFixupKind(X86::reloc_branch_4byte_pcrel)) {
355     ImmOffset -= 4;
356     // If this is a pc-relative load off _GLOBAL_OFFSET_TABLE_:
357     // leaq _GLOBAL_OFFSET_TABLE_(%rip), %r15
358     // this needs to be a GOTPC32 relocation.
359     if (StartsWithGlobalOffsetTable(Expr) != GOT_None)
360       FixupKind = MCFixupKind(X86::reloc_global_offset_table);
361   }
362   if (FixupKind == FK_PCRel_2)
363     ImmOffset -= 2;
364   if (FixupKind == FK_PCRel_1)
365     ImmOffset -= 1;
366 
367   if (ImmOffset)
368     Expr = MCBinaryExpr::createAdd(Expr, MCConstantExpr::create(ImmOffset, Ctx),
369                                    Ctx);
370 
371   // Emit a symbolic constant as a fixup and 4 zeros.
372   Fixups.push_back(MCFixup::create(CurByte, Expr, FixupKind, Loc));
373   EmitConstant(0, Size, CurByte, OS);
374 }
375 
376 void X86MCCodeEmitter::emitMemModRMByte(const MCInst &MI, unsigned Op,
377                                         unsigned RegOpcodeField,
378                                         uint64_t TSFlags, bool Rex,
379                                         unsigned &CurByte, raw_ostream &OS,
380                                         SmallVectorImpl<MCFixup> &Fixups,
381                                         const MCSubtargetInfo &STI) const {
382   const MCOperand &Disp     = MI.getOperand(Op+X86::AddrDisp);
383   const MCOperand &Base     = MI.getOperand(Op+X86::AddrBaseReg);
384   const MCOperand &Scale    = MI.getOperand(Op+X86::AddrScaleAmt);
385   const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
386   unsigned BaseReg = Base.getReg();
387   bool HasEVEX = (TSFlags & X86II::EncodingMask) == X86II::EVEX;
388 
389   // Handle %rip relative addressing.
390   if (BaseReg == X86::RIP ||
391       BaseReg == X86::EIP) {    // [disp32+rIP] in X86-64 mode
392     assert(is64BitMode(STI) && "Rip-relative addressing requires 64-bit mode");
393     assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
394     EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
395 
396     unsigned Opcode = MI.getOpcode();
397     // movq loads are handled with a special relocation form which allows the
398     // linker to eliminate some loads for GOT references which end up in the
399     // same linkage unit.
400     unsigned FixupKind = [=]() {
401       switch (Opcode) {
402       default:
403         return X86::reloc_riprel_4byte;
404       case X86::MOV64rm:
405         assert(Rex);
406         return X86::reloc_riprel_4byte_movq_load;
407       case X86::CALL64m:
408       case X86::JMP64m:
409       case X86::TAILJMPm64:
410       case X86::TEST64mr:
411       case X86::ADC64rm:
412       case X86::ADD64rm:
413       case X86::AND64rm:
414       case X86::CMP64rm:
415       case X86::OR64rm:
416       case X86::SBB64rm:
417       case X86::SUB64rm:
418       case X86::XOR64rm:
419         return Rex ? X86::reloc_riprel_4byte_relax_rex
420                    : X86::reloc_riprel_4byte_relax;
421       }
422     }();
423 
424     // rip-relative addressing is actually relative to the *next* instruction.
425     // Since an immediate can follow the mod/rm byte for an instruction, this
426     // means that we need to bias the displacement field of the instruction with
427     // the size of the immediate field. If we have this case, add it into the
428     // expression to emit.
429     // Note: rip-relative addressing using immediate displacement values should
430     // not be adjusted, assuming it was the user's intent.
431     int ImmSize = !Disp.isImm() && X86II::hasImm(TSFlags)
432                       ? X86II::getSizeOfImm(TSFlags)
433                       : 0;
434 
435     EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind),
436                   CurByte, OS, Fixups, -ImmSize);
437     return;
438   }
439 
440   unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
441 
442   // 16-bit addressing forms of the ModR/M byte have a different encoding for
443   // the R/M field and are far more limited in which registers can be used.
444   if (Is16BitMemOperand(MI, Op, STI)) {
445     if (BaseReg) {
446       // For 32-bit addressing, the row and column values in Table 2-2 are
447       // basically the same. It's AX/CX/DX/BX/SP/BP/SI/DI in that order, with
448       // some special cases. And GetX86RegNum reflects that numbering.
449       // For 16-bit addressing it's more fun, as shown in the SDM Vol 2A,
450       // Table 2-1 "16-Bit Addressing Forms with the ModR/M byte". We can only
451       // use SI/DI/BP/BX, which have "row" values 4-7 in no particular order,
452       // while values 0-3 indicate the allowed combinations (base+index) of
453       // those: 0 for BX+SI, 1 for BX+DI, 2 for BP+SI, 3 for BP+DI.
454       //
455       // R16Table[] is a lookup from the normal RegNo, to the row values from
456       // Table 2-1 for 16-bit addressing modes. Where zero means disallowed.
457       static const unsigned R16Table[] = { 0, 0, 0, 7, 0, 6, 4, 5 };
458       unsigned RMfield = R16Table[BaseRegNo];
459 
460       assert(RMfield && "invalid 16-bit base register");
461 
462       if (IndexReg.getReg()) {
463         unsigned IndexReg16 = R16Table[GetX86RegNum(IndexReg)];
464 
465         assert(IndexReg16 && "invalid 16-bit index register");
466         // We must have one of SI/DI (4,5), and one of BP/BX (6,7).
467         assert(((IndexReg16 ^ RMfield) & 2) &&
468                "invalid 16-bit base/index register combination");
469         assert(Scale.getImm() == 1 &&
470                "invalid scale for 16-bit memory reference");
471 
472         // Allow base/index to appear in either order (although GAS doesn't).
473         if (IndexReg16 & 2)
474           RMfield = (RMfield & 1) | ((7 - IndexReg16) << 1);
475         else
476           RMfield = (IndexReg16 & 1) | ((7 - RMfield) << 1);
477       }
478 
479       if (Disp.isImm() && isDisp8(Disp.getImm())) {
480         if (Disp.getImm() == 0 && RMfield != 6) {
481           // There is no displacement; just the register.
482           EmitByte(ModRMByte(0, RegOpcodeField, RMfield), CurByte, OS);
483           return;
484         }
485         // Use the [REG]+disp8 form, including for [BP] which cannot be encoded.
486         EmitByte(ModRMByte(1, RegOpcodeField, RMfield), CurByte, OS);
487         EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
488         return;
489       }
490       // This is the [REG]+disp16 case.
491       EmitByte(ModRMByte(2, RegOpcodeField, RMfield), CurByte, OS);
492     } else {
493       // There is no BaseReg; this is the plain [disp16] case.
494       EmitByte(ModRMByte(0, RegOpcodeField, 6), CurByte, OS);
495     }
496 
497     // Emit 16-bit displacement for plain disp16 or [REG]+disp16 cases.
498     EmitImmediate(Disp, MI.getLoc(), 2, FK_Data_2, CurByte, OS, Fixups);
499     return;
500   }
501 
502   // Determine whether a SIB byte is needed.
503   // If no BaseReg, issue a RIP relative instruction only if the MCE can
504   // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
505   // 2-7) and absolute references.
506 
507   if (// The SIB byte must be used if there is an index register.
508       IndexReg.getReg() == 0 &&
509       // The SIB byte must be used if the base is ESP/RSP/R12, all of which
510       // encode to an R/M value of 4, which indicates that a SIB byte is
511       // present.
512       BaseRegNo != N86::ESP &&
513       // If there is no base register and we're in 64-bit mode, we need a SIB
514       // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
515       (!is64BitMode(STI) || BaseReg != 0)) {
516 
517     if (BaseReg == 0) {          // [disp32]     in X86-32 mode
518       EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
519       EmitImmediate(Disp, MI.getLoc(), 4, FK_Data_4, CurByte, OS, Fixups);
520       return;
521     }
522 
523     // If the base is not EBP/ESP and there is no displacement, use simple
524     // indirect register encoding, this handles addresses like [EAX].  The
525     // encoding for [EBP] with no displacement means [disp32] so we handle it
526     // by emitting a displacement of 0 below.
527     if (BaseRegNo != N86::EBP) {
528       if (Disp.isImm() && Disp.getImm() == 0) {
529         EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
530         return;
531       }
532 
533       // If the displacement is @tlscall, treat it as a zero.
534       if (Disp.isExpr()) {
535         auto *Sym = dyn_cast<MCSymbolRefExpr>(Disp.getExpr());
536         if (Sym && Sym->getKind() == MCSymbolRefExpr::VK_TLSCALL) {
537           // This is exclusively used by call *a@tlscall(base). The relocation
538           // (R_386_TLSCALL or R_X86_64_TLSCALL) applies to the beginning.
539           Fixups.push_back(MCFixup::create(0, Sym, FK_NONE, MI.getLoc()));
540           EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
541           return;
542         }
543       }
544     }
545 
546     // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
547     if (Disp.isImm()) {
548       if (!HasEVEX && isDisp8(Disp.getImm())) {
549         EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
550         EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
551         return;
552       }
553       // Try EVEX compressed 8-bit displacement first; if failed, fall back to
554       // 32-bit displacement.
555       int CDisp8 = 0;
556       if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
557         EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
558         EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups,
559                       CDisp8 - Disp.getImm());
560         return;
561       }
562     }
563 
564     // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
565     EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
566     unsigned Opcode = MI.getOpcode();
567     unsigned FixupKind = Opcode == X86::MOV32rm ? X86::reloc_signed_4byte_relax
568                                                 : X86::reloc_signed_4byte;
569     EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind), CurByte, OS,
570                   Fixups);
571     return;
572   }
573 
574   // We need a SIB byte, so start by outputting the ModR/M byte first
575   assert(IndexReg.getReg() != X86::ESP &&
576          IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
577 
578   bool ForceDisp32 = false;
579   bool ForceDisp8  = false;
580   int CDisp8 = 0;
581   int ImmOffset = 0;
582   if (BaseReg == 0) {
583     // If there is no base register, we emit the special case SIB byte with
584     // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
585     EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
586     ForceDisp32 = true;
587   } else if (!Disp.isImm()) {
588     // Emit the normal disp32 encoding.
589     EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
590     ForceDisp32 = true;
591   } else if (Disp.getImm() == 0 &&
592              // Base reg can't be anything that ends up with '5' as the base
593              // reg, it is the magic [*] nomenclature that indicates no base.
594              BaseRegNo != N86::EBP) {
595     // Emit no displacement ModR/M byte
596     EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
597   } else if (!HasEVEX && isDisp8(Disp.getImm())) {
598     // Emit the disp8 encoding.
599     EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
600     ForceDisp8 = true;           // Make sure to force 8 bit disp if Base=EBP
601   } else if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
602     // Emit the disp8 encoding.
603     EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
604     ForceDisp8 = true;           // Make sure to force 8 bit disp if Base=EBP
605     ImmOffset = CDisp8 - Disp.getImm();
606   } else {
607     // Emit the normal disp32 encoding.
608     EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
609   }
610 
611   // Calculate what the SS field value should be...
612   static const unsigned SSTable[] = { ~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3 };
613   unsigned SS = SSTable[Scale.getImm()];
614 
615   if (BaseReg == 0) {
616     // Handle the SIB byte for the case where there is no base, see Intel
617     // Manual 2A, table 2-7. The displacement has already been output.
618     unsigned IndexRegNo;
619     if (IndexReg.getReg())
620       IndexRegNo = GetX86RegNum(IndexReg);
621     else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
622       IndexRegNo = 4;
623     EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
624   } else {
625     unsigned IndexRegNo;
626     if (IndexReg.getReg())
627       IndexRegNo = GetX86RegNum(IndexReg);
628     else
629       IndexRegNo = 4;   // For example [ESP+1*<noreg>+4]
630     EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
631   }
632 
633   // Do we need to output a displacement?
634   if (ForceDisp8)
635     EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups, ImmOffset);
636   else if (ForceDisp32 || Disp.getImm() != 0)
637     EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(X86::reloc_signed_4byte),
638                   CurByte, OS, Fixups);
639 }
640 
641 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
642 /// called VEX.
643 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
644                                            int MemOperand, const MCInst &MI,
645                                            const MCInstrDesc &Desc,
646                                            raw_ostream &OS) const {
647   assert(!(TSFlags & X86II::LOCK) && "Can't have LOCK VEX.");
648 
649   uint64_t Encoding = TSFlags & X86II::EncodingMask;
650   bool HasEVEX_K = TSFlags & X86II::EVEX_K;
651   bool HasVEX_4V = TSFlags & X86II::VEX_4V;
652   bool HasEVEX_RC = TSFlags & X86II::EVEX_RC;
653 
654   // VEX_R: opcode externsion equivalent to REX.R in
655   // 1's complement (inverted) form
656   //
657   //  1: Same as REX_R=0 (must be 1 in 32-bit mode)
658   //  0: Same as REX_R=1 (64 bit mode only)
659   //
660   uint8_t VEX_R = 0x1;
661   uint8_t EVEX_R2 = 0x1;
662 
663   // VEX_X: equivalent to REX.X, only used when a
664   // register is used for index in SIB Byte.
665   //
666   //  1: Same as REX.X=0 (must be 1 in 32-bit mode)
667   //  0: Same as REX.X=1 (64-bit mode only)
668   uint8_t VEX_X = 0x1;
669 
670   // VEX_B:
671   //
672   //  1: Same as REX_B=0 (ignored in 32-bit mode)
673   //  0: Same as REX_B=1 (64 bit mode only)
674   //
675   uint8_t VEX_B = 0x1;
676 
677   // VEX_W: opcode specific (use like REX.W, or used for
678   // opcode extension, or ignored, depending on the opcode byte)
679   uint8_t VEX_W = (TSFlags & X86II::VEX_W) ? 1 : 0;
680 
681   // VEX_5M (VEX m-mmmmm field):
682   //
683   //  0b00000: Reserved for future use
684   //  0b00001: implied 0F leading opcode
685   //  0b00010: implied 0F 38 leading opcode bytes
686   //  0b00011: implied 0F 3A leading opcode bytes
687   //  0b00100-0b11111: Reserved for future use
688   //  0b01000: XOP map select - 08h instructions with imm byte
689   //  0b01001: XOP map select - 09h instructions with no imm byte
690   //  0b01010: XOP map select - 0Ah instructions with imm dword
691   uint8_t VEX_5M;
692   switch (TSFlags & X86II::OpMapMask) {
693   default: llvm_unreachable("Invalid prefix!");
694   case X86II::TB:   VEX_5M = 0x1; break; // 0F
695   case X86II::T8:   VEX_5M = 0x2; break; // 0F 38
696   case X86II::TA:   VEX_5M = 0x3; break; // 0F 3A
697   case X86II::XOP8: VEX_5M = 0x8; break;
698   case X86II::XOP9: VEX_5M = 0x9; break;
699   case X86II::XOPA: VEX_5M = 0xA; break;
700   }
701 
702   // VEX_4V (VEX vvvv field): a register specifier
703   // (in 1's complement form) or 1111 if unused.
704   uint8_t VEX_4V = 0xf;
705   uint8_t EVEX_V2 = 0x1;
706 
707   // EVEX_L2/VEX_L (Vector Length):
708   //
709   // L2 L
710   //  0 0: scalar or 128-bit vector
711   //  0 1: 256-bit vector
712   //  1 0: 512-bit vector
713   //
714   uint8_t VEX_L = (TSFlags & X86II::VEX_L) ? 1 : 0;
715   uint8_t EVEX_L2 = (TSFlags & X86II::EVEX_L2) ? 1 : 0;
716 
717   // VEX_PP: opcode extension providing equivalent
718   // functionality of a SIMD prefix
719   //
720   //  0b00: None
721   //  0b01: 66
722   //  0b10: F3
723   //  0b11: F2
724   //
725   uint8_t VEX_PP = 0;
726   switch (TSFlags & X86II::OpPrefixMask) {
727   case X86II::PD: VEX_PP = 0x1; break; // 66
728   case X86II::XS: VEX_PP = 0x2; break; // F3
729   case X86II::XD: VEX_PP = 0x3; break; // F2
730   }
731 
732   // EVEX_U
733   uint8_t EVEX_U = 1; // Always '1' so far
734 
735   // EVEX_z
736   uint8_t EVEX_z = (HasEVEX_K && (TSFlags & X86II::EVEX_Z)) ? 1 : 0;
737 
738   // EVEX_b
739   uint8_t EVEX_b = (TSFlags & X86II::EVEX_B) ? 1 : 0;
740 
741   // EVEX_rc
742   uint8_t EVEX_rc = 0;
743 
744   // EVEX_aaa
745   uint8_t EVEX_aaa = 0;
746 
747   bool EncodeRC = false;
748 
749   // Classify VEX_B, VEX_4V, VEX_R, VEX_X
750   unsigned NumOps = Desc.getNumOperands();
751   unsigned CurOp = X86II::getOperandBias(Desc);
752 
753   switch (TSFlags & X86II::FormMask) {
754   default: llvm_unreachable("Unexpected form in EmitVEXOpcodePrefix!");
755   case X86II::RawFrm:
756     break;
757   case X86II::MRMDestMem: {
758     // MRMDestMem instructions forms:
759     //  MemAddr, src1(ModR/M)
760     //  MemAddr, src1(VEX_4V), src2(ModR/M)
761     //  MemAddr, src1(ModR/M), imm8
762     //
763     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
764     VEX_B = ~(BaseRegEnc >> 3) & 1;
765     unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
766     VEX_X = ~(IndexRegEnc >> 3) & 1;
767     if (!HasVEX_4V) // Only needed with VSIB which don't use VVVV.
768       EVEX_V2 = ~(IndexRegEnc >> 4) & 1;
769 
770     CurOp += X86::AddrNumOperands;
771 
772     if (HasEVEX_K)
773       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
774 
775     if (HasVEX_4V) {
776       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
777       VEX_4V = ~VRegEnc & 0xf;
778       EVEX_V2 = ~(VRegEnc >> 4) & 1;
779     }
780 
781     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
782     VEX_R = ~(RegEnc >> 3) & 1;
783     EVEX_R2 = ~(RegEnc >> 4) & 1;
784     break;
785   }
786   case X86II::MRMSrcMem: {
787     // MRMSrcMem instructions forms:
788     //  src1(ModR/M), MemAddr
789     //  src1(ModR/M), src2(VEX_4V), MemAddr
790     //  src1(ModR/M), MemAddr, imm8
791     //  src1(ModR/M), MemAddr, src2(Imm[7:4])
792     //
793     //  FMA4:
794     //  dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(Imm[7:4])
795     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
796     VEX_R = ~(RegEnc >> 3) & 1;
797     EVEX_R2 = ~(RegEnc >> 4) & 1;
798 
799     if (HasEVEX_K)
800       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
801 
802     if (HasVEX_4V) {
803       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
804       VEX_4V = ~VRegEnc & 0xf;
805       EVEX_V2 = ~(VRegEnc >> 4) & 1;
806     }
807 
808     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
809     VEX_B = ~(BaseRegEnc >> 3) & 1;
810     unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
811     VEX_X = ~(IndexRegEnc >> 3) & 1;
812     if (!HasVEX_4V) // Only needed with VSIB which don't use VVVV.
813       EVEX_V2 = ~(IndexRegEnc >> 4) & 1;
814 
815     break;
816   }
817   case X86II::MRMSrcMem4VOp3: {
818     // Instruction format for 4VOp3:
819     //   src1(ModR/M), MemAddr, src3(VEX_4V)
820     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
821     VEX_R = ~(RegEnc >> 3) & 1;
822 
823     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
824     VEX_B = ~(BaseRegEnc >> 3) & 1;
825     unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
826     VEX_X = ~(IndexRegEnc >> 3) & 1;
827 
828     VEX_4V = ~getX86RegEncoding(MI, CurOp + X86::AddrNumOperands) & 0xf;
829     break;
830   }
831   case X86II::MRMSrcMemOp4: {
832     //  dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
833     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
834     VEX_R = ~(RegEnc >> 3) & 1;
835 
836     unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
837     VEX_4V = ~VRegEnc & 0xf;
838 
839     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
840     VEX_B = ~(BaseRegEnc >> 3) & 1;
841     unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
842     VEX_X = ~(IndexRegEnc >> 3) & 1;
843     break;
844   }
845   case X86II::MRM0m: case X86II::MRM1m:
846   case X86II::MRM2m: case X86II::MRM3m:
847   case X86II::MRM4m: case X86II::MRM5m:
848   case X86II::MRM6m: case X86II::MRM7m: {
849     // MRM[0-9]m instructions forms:
850     //  MemAddr
851     //  src1(VEX_4V), MemAddr
852     if (HasVEX_4V) {
853       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
854       VEX_4V = ~VRegEnc & 0xf;
855       EVEX_V2 = ~(VRegEnc >> 4) & 1;
856     }
857 
858     if (HasEVEX_K)
859       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
860 
861     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
862     VEX_B = ~(BaseRegEnc >> 3) & 1;
863     unsigned IndexRegEnc = getX86RegEncoding(MI, MemOperand+X86::AddrIndexReg);
864     VEX_X = ~(IndexRegEnc >> 3) & 1;
865     if (!HasVEX_4V) // Only needed with VSIB which don't use VVVV.
866       EVEX_V2 = ~(IndexRegEnc >> 4) & 1;
867 
868     break;
869   }
870   case X86II::MRMSrcReg: {
871     // MRMSrcReg instructions forms:
872     //  dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(Imm[7:4])
873     //  dst(ModR/M), src1(ModR/M)
874     //  dst(ModR/M), src1(ModR/M), imm8
875     //
876     //  FMA4:
877     //  dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
878     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
879     VEX_R = ~(RegEnc >> 3) & 1;
880     EVEX_R2 = ~(RegEnc >> 4) & 1;
881 
882     if (HasEVEX_K)
883       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
884 
885     if (HasVEX_4V) {
886       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
887       VEX_4V = ~VRegEnc & 0xf;
888       EVEX_V2 = ~(VRegEnc >> 4) & 1;
889     }
890 
891     RegEnc = getX86RegEncoding(MI, CurOp++);
892     VEX_B = ~(RegEnc >> 3) & 1;
893     VEX_X = ~(RegEnc >> 4) & 1;
894 
895     if (EVEX_b) {
896       if (HasEVEX_RC) {
897         unsigned RcOperand = NumOps-1;
898         assert(RcOperand >= CurOp);
899         EVEX_rc = MI.getOperand(RcOperand).getImm();
900         assert(EVEX_rc <= 3 && "Invalid rounding control!");
901       }
902       EncodeRC = true;
903     }
904     break;
905   }
906   case X86II::MRMSrcReg4VOp3: {
907     // Instruction format for 4VOp3:
908     //   src1(ModR/M), src2(ModR/M), src3(VEX_4V)
909     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
910     VEX_R = ~(RegEnc >> 3) & 1;
911 
912     RegEnc = getX86RegEncoding(MI, CurOp++);
913     VEX_B = ~(RegEnc >> 3) & 1;
914 
915     VEX_4V = ~getX86RegEncoding(MI, CurOp++) & 0xf;
916     break;
917   }
918   case X86II::MRMSrcRegOp4: {
919     //  dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
920     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
921     VEX_R = ~(RegEnc >> 3) & 1;
922 
923     unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
924     VEX_4V = ~VRegEnc & 0xf;
925 
926     // Skip second register source (encoded in Imm[7:4])
927     ++CurOp;
928 
929     RegEnc = getX86RegEncoding(MI, CurOp++);
930     VEX_B = ~(RegEnc >> 3) & 1;
931     VEX_X = ~(RegEnc >> 4) & 1;
932     break;
933   }
934   case X86II::MRMDestReg: {
935     // MRMDestReg instructions forms:
936     //  dst(ModR/M), src(ModR/M)
937     //  dst(ModR/M), src(ModR/M), imm8
938     //  dst(ModR/M), src1(VEX_4V), src2(ModR/M)
939     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
940     VEX_B = ~(RegEnc >> 3) & 1;
941     VEX_X = ~(RegEnc >> 4) & 1;
942 
943     if (HasEVEX_K)
944       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
945 
946     if (HasVEX_4V) {
947       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
948       VEX_4V = ~VRegEnc & 0xf;
949       EVEX_V2 = ~(VRegEnc >> 4) & 1;
950     }
951 
952     RegEnc = getX86RegEncoding(MI, CurOp++);
953     VEX_R = ~(RegEnc >> 3) & 1;
954     EVEX_R2 = ~(RegEnc >> 4) & 1;
955     if (EVEX_b)
956       EncodeRC = true;
957     break;
958   }
959   case X86II::MRM0r: case X86II::MRM1r:
960   case X86II::MRM2r: case X86II::MRM3r:
961   case X86II::MRM4r: case X86II::MRM5r:
962   case X86II::MRM6r: case X86II::MRM7r: {
963     // MRM0r-MRM7r instructions forms:
964     //  dst(VEX_4V), src(ModR/M), imm8
965     if (HasVEX_4V) {
966       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
967       VEX_4V = ~VRegEnc & 0xf;
968       EVEX_V2 = ~(VRegEnc >> 4) & 1;
969     }
970     if (HasEVEX_K)
971       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
972 
973     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
974     VEX_B = ~(RegEnc >> 3) & 1;
975     VEX_X = ~(RegEnc >> 4) & 1;
976     break;
977   }
978   }
979 
980   if (Encoding == X86II::VEX || Encoding == X86II::XOP) {
981     // VEX opcode prefix can have 2 or 3 bytes
982     //
983     //  3 bytes:
984     //    +-----+ +--------------+ +-------------------+
985     //    | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
986     //    +-----+ +--------------+ +-------------------+
987     //  2 bytes:
988     //    +-----+ +-------------------+
989     //    | C5h | | R | vvvv | L | pp |
990     //    +-----+ +-------------------+
991     //
992     //  XOP uses a similar prefix:
993     //    +-----+ +--------------+ +-------------------+
994     //    | 8Fh | | RXB | m-mmmm | | W | vvvv | L | pp |
995     //    +-----+ +--------------+ +-------------------+
996     uint8_t LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
997 
998     // Can we use the 2 byte VEX prefix?
999     if (!(MI.getFlags() & X86::IP_USE_VEX3) &&
1000         Encoding == X86II::VEX && VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) {
1001       EmitByte(0xC5, CurByte, OS);
1002       EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
1003       return;
1004     }
1005 
1006     // 3 byte VEX prefix
1007     EmitByte(Encoding == X86II::XOP ? 0x8F : 0xC4, CurByte, OS);
1008     EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
1009     EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
1010   } else {
1011     assert(Encoding == X86II::EVEX && "unknown encoding!");
1012     // EVEX opcode prefix can have 4 bytes
1013     //
1014     // +-----+ +--------------+ +-------------------+ +------------------------+
1015     // | 62h | | RXBR' | 00mm | | W | vvvv | U | pp | | z | L'L | b | v' | aaa |
1016     // +-----+ +--------------+ +-------------------+ +------------------------+
1017     assert((VEX_5M & 0x3) == VEX_5M
1018            && "More than 2 significant bits in VEX.m-mmmm fields for EVEX!");
1019 
1020     EmitByte(0x62, CurByte, OS);
1021     EmitByte((VEX_R   << 7) |
1022              (VEX_X   << 6) |
1023              (VEX_B   << 5) |
1024              (EVEX_R2 << 4) |
1025              VEX_5M, CurByte, OS);
1026     EmitByte((VEX_W   << 7) |
1027              (VEX_4V  << 3) |
1028              (EVEX_U  << 2) |
1029              VEX_PP, CurByte, OS);
1030     if (EncodeRC)
1031       EmitByte((EVEX_z  << 7) |
1032                (EVEX_rc << 5) |
1033                (EVEX_b  << 4) |
1034                (EVEX_V2 << 3) |
1035                EVEX_aaa, CurByte, OS);
1036     else
1037       EmitByte((EVEX_z  << 7) |
1038                (EVEX_L2 << 6) |
1039                (VEX_L   << 5) |
1040                (EVEX_b  << 4) |
1041                (EVEX_V2 << 3) |
1042                EVEX_aaa, CurByte, OS);
1043   }
1044 }
1045 
1046 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
1047 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
1048 /// size, and 3) use of X86-64 extended registers.
1049 uint8_t X86MCCodeEmitter::DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
1050                                              int MemOperand,
1051                                              const MCInstrDesc &Desc) const {
1052   uint8_t REX = 0;
1053   bool UsesHighByteReg = false;
1054 
1055   if (TSFlags & X86II::REX_W)
1056     REX |= 1 << 3; // set REX.W
1057 
1058   if (MI.getNumOperands() == 0) return REX;
1059 
1060   unsigned NumOps = MI.getNumOperands();
1061   unsigned CurOp = X86II::getOperandBias(Desc);
1062 
1063   // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
1064   for (unsigned i = CurOp; i != NumOps; ++i) {
1065     const MCOperand &MO = MI.getOperand(i);
1066     if (!MO.isReg()) continue;
1067     unsigned Reg = MO.getReg();
1068     if (Reg == X86::AH || Reg == X86::BH || Reg == X86::CH || Reg == X86::DH)
1069       UsesHighByteReg = true;
1070     if (X86II::isX86_64NonExtLowByteReg(Reg))
1071       // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
1072       // that returns non-zero.
1073       REX |= 0x40; // REX fixed encoding prefix
1074   }
1075 
1076   switch (TSFlags & X86II::FormMask) {
1077   case X86II::AddRegFrm:
1078     REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1079     break;
1080   case X86II::MRMSrcReg:
1081   case X86II::MRMSrcRegCC:
1082     REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1083     REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1084     break;
1085   case X86II::MRMSrcMem:
1086   case X86II::MRMSrcMemCC:
1087     REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1088     REX |= isREXExtendedReg(MI, MemOperand+X86::AddrBaseReg) << 0; // REX.B
1089     REX |= isREXExtendedReg(MI, MemOperand+X86::AddrIndexReg) << 1; // REX.X
1090     CurOp += X86::AddrNumOperands;
1091     break;
1092   case X86II::MRMDestReg:
1093     REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1094     REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1095     break;
1096   case X86II::MRMDestMem:
1097     REX |= isREXExtendedReg(MI, MemOperand+X86::AddrBaseReg) << 0; // REX.B
1098     REX |= isREXExtendedReg(MI, MemOperand+X86::AddrIndexReg) << 1; // REX.X
1099     CurOp += X86::AddrNumOperands;
1100     REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1101     break;
1102   case X86II::MRMXmCC: case X86II::MRMXm:
1103   case X86II::MRM0m: case X86II::MRM1m:
1104   case X86II::MRM2m: case X86II::MRM3m:
1105   case X86II::MRM4m: case X86II::MRM5m:
1106   case X86II::MRM6m: case X86II::MRM7m:
1107     REX |= isREXExtendedReg(MI, MemOperand+X86::AddrBaseReg) << 0; // REX.B
1108     REX |= isREXExtendedReg(MI, MemOperand+X86::AddrIndexReg) << 1; // REX.X
1109     break;
1110   case X86II::MRMXrCC: case X86II::MRMXr:
1111   case X86II::MRM0r: case X86II::MRM1r:
1112   case X86II::MRM2r: case X86II::MRM3r:
1113   case X86II::MRM4r: case X86II::MRM5r:
1114   case X86II::MRM6r: case X86II::MRM7r:
1115     REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1116     break;
1117   }
1118   if (REX && UsesHighByteReg)
1119     report_fatal_error("Cannot encode high byte register in REX-prefixed instruction");
1120 
1121   return REX;
1122 }
1123 
1124 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
1125 void X86MCCodeEmitter::EmitSegmentOverridePrefix(unsigned &CurByte,
1126                                                  unsigned SegOperand,
1127                                                  const MCInst &MI,
1128                                                  raw_ostream &OS) const {
1129   // Check for explicit segment override on memory operand.
1130   switch (MI.getOperand(SegOperand).getReg()) {
1131   default: llvm_unreachable("Unknown segment register!");
1132   case 0: break;
1133   case X86::CS: EmitByte(0x2E, CurByte, OS); break;
1134   case X86::SS: EmitByte(0x36, CurByte, OS); break;
1135   case X86::DS: EmitByte(0x3E, CurByte, OS); break;
1136   case X86::ES: EmitByte(0x26, CurByte, OS); break;
1137   case X86::FS: EmitByte(0x64, CurByte, OS); break;
1138   case X86::GS: EmitByte(0x65, CurByte, OS); break;
1139   }
1140 }
1141 
1142 /// Emit all instruction prefixes prior to the opcode.
1143 ///
1144 /// MemOperand is the operand # of the start of a memory operand if present.  If
1145 /// Not present, it is -1.
1146 ///
1147 /// Returns true if a REX prefix was used.
1148 bool X86MCCodeEmitter::emitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
1149                                         int MemOperand, const MCInst &MI,
1150                                         const MCInstrDesc &Desc,
1151                                         const MCSubtargetInfo &STI,
1152                                         raw_ostream &OS) const {
1153   bool Ret = false;
1154   // Emit the operand size opcode prefix as needed.
1155   if ((TSFlags & X86II::OpSizeMask) == (is16BitMode(STI) ? X86II::OpSize32
1156                                                          : X86II::OpSize16))
1157     EmitByte(0x66, CurByte, OS);
1158 
1159   // Emit the LOCK opcode prefix.
1160   if (TSFlags & X86II::LOCK || MI.getFlags() & X86::IP_HAS_LOCK)
1161     EmitByte(0xF0, CurByte, OS);
1162 
1163   // Emit the NOTRACK opcode prefix.
1164   if (TSFlags & X86II::NOTRACK || MI.getFlags() & X86::IP_HAS_NOTRACK)
1165     EmitByte(0x3E, CurByte, OS);
1166 
1167   switch (TSFlags & X86II::OpPrefixMask) {
1168   case X86II::PD:   // 66
1169     EmitByte(0x66, CurByte, OS);
1170     break;
1171   case X86II::XS:   // F3
1172     EmitByte(0xF3, CurByte, OS);
1173     break;
1174   case X86II::XD:   // F2
1175     EmitByte(0xF2, CurByte, OS);
1176     break;
1177   }
1178 
1179   // Handle REX prefix.
1180   // FIXME: Can this come before F2 etc to simplify emission?
1181   if (is64BitMode(STI)) {
1182     if (uint8_t REX = DetermineREXPrefix(MI, TSFlags, MemOperand, Desc)) {
1183       EmitByte(0x40 | REX, CurByte, OS);
1184       Ret = true;
1185     }
1186   } else {
1187     assert(!(TSFlags & X86II::REX_W) && "REX.W requires 64bit mode.");
1188   }
1189 
1190   // 0x0F escape code must be emitted just before the opcode.
1191   switch (TSFlags & X86II::OpMapMask) {
1192   case X86II::TB:         // Two-byte opcode map
1193   case X86II::T8:         // 0F 38
1194   case X86II::TA:         // 0F 3A
1195   case X86II::ThreeDNow:  // 0F 0F, second 0F emitted by caller.
1196     EmitByte(0x0F, CurByte, OS);
1197     break;
1198   }
1199 
1200   switch (TSFlags & X86II::OpMapMask) {
1201   case X86II::T8:    // 0F 38
1202     EmitByte(0x38, CurByte, OS);
1203     break;
1204   case X86II::TA:    // 0F 3A
1205     EmitByte(0x3A, CurByte, OS);
1206     break;
1207   }
1208   return Ret;
1209 }
1210 
1211 void X86MCCodeEmitter::
1212 encodeInstruction(const MCInst &MI, raw_ostream &OS,
1213                   SmallVectorImpl<MCFixup> &Fixups,
1214                   const MCSubtargetInfo &STI) const {
1215   unsigned Opcode = MI.getOpcode();
1216   const MCInstrDesc &Desc = MCII.get(Opcode);
1217   uint64_t TSFlags = Desc.TSFlags;
1218   unsigned Flags = MI.getFlags();
1219 
1220   // Pseudo instructions don't get encoded.
1221   if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
1222     return;
1223 
1224   unsigned NumOps = Desc.getNumOperands();
1225   unsigned CurOp = X86II::getOperandBias(Desc);
1226 
1227   // Keep track of the current byte being emitted.
1228   unsigned CurByte = 0;
1229 
1230   // Encoding type for this instruction.
1231   uint64_t Encoding = TSFlags & X86II::EncodingMask;
1232 
1233   // It uses the VEX.VVVV field?
1234   bool HasVEX_4V = TSFlags & X86II::VEX_4V;
1235   bool HasVEX_I8Reg = (TSFlags & X86II::ImmMask) == X86II::Imm8Reg;
1236 
1237   // It uses the EVEX.aaa field?
1238   bool HasEVEX_K = TSFlags & X86II::EVEX_K;
1239   bool HasEVEX_RC = TSFlags & X86II::EVEX_RC;
1240 
1241   // Used if a register is encoded in 7:4 of immediate.
1242   unsigned I8RegNum = 0;
1243 
1244   // Determine where the memory operand starts, if present.
1245   int MemoryOperand = X86II::getMemoryOperandNo(TSFlags);
1246   if (MemoryOperand != -1) MemoryOperand += CurOp;
1247 
1248   // Emit segment override opcode prefix as needed.
1249   if (MemoryOperand >= 0)
1250     EmitSegmentOverridePrefix(CurByte, MemoryOperand+X86::AddrSegmentReg,
1251                               MI, OS);
1252 
1253   // Emit the repeat opcode prefix as needed.
1254   if (TSFlags & X86II::REP || Flags & X86::IP_HAS_REPEAT)
1255     EmitByte(0xF3, CurByte, OS);
1256   if (Flags & X86::IP_HAS_REPEAT_NE)
1257     EmitByte(0xF2, CurByte, OS);
1258 
1259   // Emit the address size opcode prefix as needed.
1260   bool need_address_override;
1261   uint64_t AdSize = TSFlags & X86II::AdSizeMask;
1262   if ((is16BitMode(STI) && AdSize == X86II::AdSize32) ||
1263       (is32BitMode(STI) && AdSize == X86II::AdSize16) ||
1264       (is64BitMode(STI) && AdSize == X86II::AdSize32)) {
1265     need_address_override = true;
1266   } else if (MemoryOperand < 0) {
1267     need_address_override = false;
1268   } else if (is64BitMode(STI)) {
1269     assert(!Is16BitMemOperand(MI, MemoryOperand, STI));
1270     need_address_override = Is32BitMemOperand(MI, MemoryOperand);
1271   } else if (is32BitMode(STI)) {
1272     assert(!Is64BitMemOperand(MI, MemoryOperand));
1273     need_address_override = Is16BitMemOperand(MI, MemoryOperand, STI);
1274   } else {
1275     assert(is16BitMode(STI));
1276     assert(!Is64BitMemOperand(MI, MemoryOperand));
1277     need_address_override = !Is16BitMemOperand(MI, MemoryOperand, STI);
1278   }
1279 
1280   if (need_address_override)
1281     EmitByte(0x67, CurByte, OS);
1282 
1283   bool Rex = false;
1284   if (Encoding == 0)
1285     Rex = emitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, STI, OS);
1286   else
1287     EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
1288 
1289   uint8_t BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
1290 
1291   if ((TSFlags & X86II::OpMapMask) == X86II::ThreeDNow)
1292     BaseOpcode = 0x0F;   // Weird 3DNow! encoding.
1293 
1294   unsigned OpcodeOffset = 0;
1295 
1296   uint64_t Form = TSFlags & X86II::FormMask;
1297   switch (Form) {
1298   default: errs() << "FORM: " << Form << "\n";
1299     llvm_unreachable("Unknown FormMask value in X86MCCodeEmitter!");
1300   case X86II::Pseudo:
1301     llvm_unreachable("Pseudo instruction shouldn't be emitted");
1302   case X86II::RawFrmDstSrc: {
1303     unsigned siReg = MI.getOperand(1).getReg();
1304     assert(((siReg == X86::SI && MI.getOperand(0).getReg() == X86::DI) ||
1305             (siReg == X86::ESI && MI.getOperand(0).getReg() == X86::EDI) ||
1306             (siReg == X86::RSI && MI.getOperand(0).getReg() == X86::RDI)) &&
1307            "SI and DI register sizes do not match");
1308     // Emit segment override opcode prefix as needed (not for %ds).
1309     if (MI.getOperand(2).getReg() != X86::DS)
1310       EmitSegmentOverridePrefix(CurByte, 2, MI, OS);
1311     // Emit AdSize prefix as needed.
1312     if ((!is32BitMode(STI) && siReg == X86::ESI) ||
1313         (is32BitMode(STI) && siReg == X86::SI))
1314       EmitByte(0x67, CurByte, OS);
1315     CurOp += 3; // Consume operands.
1316     EmitByte(BaseOpcode, CurByte, OS);
1317     break;
1318   }
1319   case X86II::RawFrmSrc: {
1320     unsigned siReg = MI.getOperand(0).getReg();
1321     // Emit segment override opcode prefix as needed (not for %ds).
1322     if (MI.getOperand(1).getReg() != X86::DS)
1323       EmitSegmentOverridePrefix(CurByte, 1, MI, OS);
1324     // Emit AdSize prefix as needed.
1325     if ((!is32BitMode(STI) && siReg == X86::ESI) ||
1326         (is32BitMode(STI) && siReg == X86::SI))
1327       EmitByte(0x67, CurByte, OS);
1328     CurOp += 2; // Consume operands.
1329     EmitByte(BaseOpcode, CurByte, OS);
1330     break;
1331   }
1332   case X86II::RawFrmDst: {
1333     unsigned siReg = MI.getOperand(0).getReg();
1334     // Emit AdSize prefix as needed.
1335     if ((!is32BitMode(STI) && siReg == X86::EDI) ||
1336         (is32BitMode(STI) && siReg == X86::DI))
1337       EmitByte(0x67, CurByte, OS);
1338     ++CurOp; // Consume operand.
1339     EmitByte(BaseOpcode, CurByte, OS);
1340     break;
1341   }
1342   case X86II::AddCCFrm: {
1343     // This will be added to the opcode in the fallthrough.
1344     OpcodeOffset = MI.getOperand(NumOps - 1).getImm();
1345     assert(OpcodeOffset < 16 && "Unexpected opcode offset!");
1346     --NumOps; // Drop the operand from the end.
1347     LLVM_FALLTHROUGH;
1348   case X86II::RawFrm:
1349     EmitByte(BaseOpcode + OpcodeOffset, CurByte, OS);
1350 
1351     if (!is64BitMode(STI) || !isPCRel32Branch(MI))
1352       break;
1353 
1354     const MCOperand &Op = MI.getOperand(CurOp++);
1355     EmitImmediate(Op, MI.getLoc(), X86II::getSizeOfImm(TSFlags),
1356                   MCFixupKind(X86::reloc_branch_4byte_pcrel), CurByte, OS,
1357                   Fixups);
1358     break;
1359   }
1360   case X86II::RawFrmMemOffs:
1361     // Emit segment override opcode prefix as needed.
1362     EmitSegmentOverridePrefix(CurByte, 1, MI, OS);
1363     EmitByte(BaseOpcode, CurByte, OS);
1364     EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1365                   X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1366                   CurByte, OS, Fixups);
1367     ++CurOp; // skip segment operand
1368     break;
1369   case X86II::RawFrmImm8:
1370     EmitByte(BaseOpcode, CurByte, OS);
1371     EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1372                   X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1373                   CurByte, OS, Fixups);
1374     EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 1, FK_Data_1, CurByte,
1375                   OS, Fixups);
1376     break;
1377   case X86II::RawFrmImm16:
1378     EmitByte(BaseOpcode, CurByte, OS);
1379     EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1380                   X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1381                   CurByte, OS, Fixups);
1382     EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 2, FK_Data_2, CurByte,
1383                   OS, Fixups);
1384     break;
1385 
1386   case X86II::AddRegFrm:
1387     EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
1388     break;
1389 
1390   case X86II::MRMDestReg: {
1391     EmitByte(BaseOpcode, CurByte, OS);
1392     unsigned SrcRegNum = CurOp + 1;
1393 
1394     if (HasEVEX_K) // Skip writemask
1395       ++SrcRegNum;
1396 
1397     if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1398       ++SrcRegNum;
1399 
1400     EmitRegModRMByte(MI.getOperand(CurOp),
1401                      GetX86RegNum(MI.getOperand(SrcRegNum)), CurByte, OS);
1402     CurOp = SrcRegNum + 1;
1403     break;
1404   }
1405   case X86II::MRMDestMem: {
1406     EmitByte(BaseOpcode, CurByte, OS);
1407     unsigned SrcRegNum = CurOp + X86::AddrNumOperands;
1408 
1409     if (HasEVEX_K) // Skip writemask
1410       ++SrcRegNum;
1411 
1412     if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1413       ++SrcRegNum;
1414 
1415     emitMemModRMByte(MI, CurOp, GetX86RegNum(MI.getOperand(SrcRegNum)), TSFlags,
1416                      Rex, CurByte, OS, Fixups, STI);
1417     CurOp = SrcRegNum + 1;
1418     break;
1419   }
1420   case X86II::MRMSrcReg: {
1421     EmitByte(BaseOpcode, CurByte, OS);
1422     unsigned SrcRegNum = CurOp + 1;
1423 
1424     if (HasEVEX_K) // Skip writemask
1425       ++SrcRegNum;
1426 
1427     if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1428       ++SrcRegNum;
1429 
1430     EmitRegModRMByte(MI.getOperand(SrcRegNum),
1431                      GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1432     CurOp = SrcRegNum + 1;
1433     if (HasVEX_I8Reg)
1434       I8RegNum = getX86RegEncoding(MI, CurOp++);
1435     // do not count the rounding control operand
1436     if (HasEVEX_RC)
1437       --NumOps;
1438     break;
1439   }
1440   case X86II::MRMSrcReg4VOp3: {
1441     EmitByte(BaseOpcode, CurByte, OS);
1442     unsigned SrcRegNum = CurOp + 1;
1443 
1444     EmitRegModRMByte(MI.getOperand(SrcRegNum),
1445                      GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1446     CurOp = SrcRegNum + 1;
1447     ++CurOp; // Encoded in VEX.VVVV
1448     break;
1449   }
1450   case X86II::MRMSrcRegOp4: {
1451     EmitByte(BaseOpcode, CurByte, OS);
1452     unsigned SrcRegNum = CurOp + 1;
1453 
1454     // Skip 1st src (which is encoded in VEX_VVVV)
1455     ++SrcRegNum;
1456 
1457     // Capture 2nd src (which is encoded in Imm[7:4])
1458     assert(HasVEX_I8Reg && "MRMSrcRegOp4 should imply VEX_I8Reg");
1459     I8RegNum = getX86RegEncoding(MI, SrcRegNum++);
1460 
1461     EmitRegModRMByte(MI.getOperand(SrcRegNum),
1462                      GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1463     CurOp = SrcRegNum + 1;
1464     break;
1465   }
1466   case X86II::MRMSrcRegCC: {
1467     unsigned FirstOp = CurOp++;
1468     unsigned SecondOp = CurOp++;
1469 
1470     unsigned CC = MI.getOperand(CurOp++).getImm();
1471     EmitByte(BaseOpcode + CC, CurByte, OS);
1472 
1473     EmitRegModRMByte(MI.getOperand(SecondOp),
1474                      GetX86RegNum(MI.getOperand(FirstOp)), CurByte, OS);
1475     break;
1476   }
1477   case X86II::MRMSrcMem: {
1478     unsigned FirstMemOp = CurOp+1;
1479 
1480     if (HasEVEX_K) // Skip writemask
1481       ++FirstMemOp;
1482 
1483     if (HasVEX_4V)
1484       ++FirstMemOp;  // Skip the register source (which is encoded in VEX_VVVV).
1485 
1486     EmitByte(BaseOpcode, CurByte, OS);
1487 
1488     emitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1489                      TSFlags, Rex, CurByte, OS, Fixups, STI);
1490     CurOp = FirstMemOp + X86::AddrNumOperands;
1491     if (HasVEX_I8Reg)
1492       I8RegNum = getX86RegEncoding(MI, CurOp++);
1493     break;
1494   }
1495   case X86II::MRMSrcMem4VOp3: {
1496     unsigned FirstMemOp = CurOp+1;
1497 
1498     EmitByte(BaseOpcode, CurByte, OS);
1499 
1500     emitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1501                      TSFlags, Rex, CurByte, OS, Fixups, STI);
1502     CurOp = FirstMemOp + X86::AddrNumOperands;
1503     ++CurOp; // Encoded in VEX.VVVV.
1504     break;
1505   }
1506   case X86II::MRMSrcMemOp4: {
1507     unsigned FirstMemOp = CurOp+1;
1508 
1509     ++FirstMemOp;  // Skip the register source (which is encoded in VEX_VVVV).
1510 
1511     // Capture second register source (encoded in Imm[7:4])
1512     assert(HasVEX_I8Reg && "MRMSrcRegOp4 should imply VEX_I8Reg");
1513     I8RegNum = getX86RegEncoding(MI, FirstMemOp++);
1514 
1515     EmitByte(BaseOpcode, CurByte, OS);
1516 
1517     emitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1518                      TSFlags, Rex, CurByte, OS, Fixups, STI);
1519     CurOp = FirstMemOp + X86::AddrNumOperands;
1520     break;
1521   }
1522   case X86II::MRMSrcMemCC: {
1523     unsigned RegOp = CurOp++;
1524     unsigned FirstMemOp = CurOp;
1525     CurOp = FirstMemOp + X86::AddrNumOperands;
1526 
1527     unsigned CC = MI.getOperand(CurOp++).getImm();
1528     EmitByte(BaseOpcode + CC, CurByte, OS);
1529 
1530     emitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(RegOp)),
1531                      TSFlags, Rex, CurByte, OS, Fixups, STI);
1532     break;
1533   }
1534 
1535   case X86II::MRMXrCC: {
1536     unsigned RegOp = CurOp++;
1537 
1538     unsigned CC = MI.getOperand(CurOp++).getImm();
1539     EmitByte(BaseOpcode + CC, CurByte, OS);
1540     EmitRegModRMByte(MI.getOperand(RegOp), 0, CurByte, OS);
1541     break;
1542   }
1543 
1544   case X86II::MRMXr:
1545   case X86II::MRM0r: case X86II::MRM1r:
1546   case X86II::MRM2r: case X86II::MRM3r:
1547   case X86II::MRM4r: case X86II::MRM5r:
1548   case X86II::MRM6r: case X86II::MRM7r:
1549     if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1550       ++CurOp;
1551     if (HasEVEX_K) // Skip writemask
1552       ++CurOp;
1553     EmitByte(BaseOpcode, CurByte, OS);
1554     EmitRegModRMByte(MI.getOperand(CurOp++),
1555                      (Form == X86II::MRMXr) ? 0 : Form-X86II::MRM0r,
1556                      CurByte, OS);
1557     break;
1558 
1559   case X86II::MRMXmCC: {
1560     unsigned FirstMemOp = CurOp;
1561     CurOp = FirstMemOp + X86::AddrNumOperands;
1562 
1563     unsigned CC = MI.getOperand(CurOp++).getImm();
1564     EmitByte(BaseOpcode + CC, CurByte, OS);
1565 
1566     emitMemModRMByte(MI, FirstMemOp, 0, TSFlags, Rex, CurByte, OS, Fixups, STI);
1567     break;
1568   }
1569 
1570   case X86II::MRMXm:
1571   case X86II::MRM0m: case X86II::MRM1m:
1572   case X86II::MRM2m: case X86II::MRM3m:
1573   case X86II::MRM4m: case X86II::MRM5m:
1574   case X86II::MRM6m: case X86II::MRM7m:
1575     if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1576       ++CurOp;
1577     if (HasEVEX_K) // Skip writemask
1578       ++CurOp;
1579     EmitByte(BaseOpcode, CurByte, OS);
1580     emitMemModRMByte(MI, CurOp,
1581                      (Form == X86II::MRMXm) ? 0 : Form - X86II::MRM0m, TSFlags,
1582                      Rex, CurByte, OS, Fixups, STI);
1583     CurOp += X86::AddrNumOperands;
1584     break;
1585 
1586   case X86II::MRM_C0: case X86II::MRM_C1: case X86II::MRM_C2:
1587   case X86II::MRM_C3: case X86II::MRM_C4: case X86II::MRM_C5:
1588   case X86II::MRM_C6: case X86II::MRM_C7: case X86II::MRM_C8:
1589   case X86II::MRM_C9: case X86II::MRM_CA: case X86II::MRM_CB:
1590   case X86II::MRM_CC: case X86II::MRM_CD: case X86II::MRM_CE:
1591   case X86II::MRM_CF: case X86II::MRM_D0: case X86II::MRM_D1:
1592   case X86II::MRM_D2: case X86II::MRM_D3: case X86II::MRM_D4:
1593   case X86II::MRM_D5: case X86II::MRM_D6: case X86II::MRM_D7:
1594   case X86II::MRM_D8: case X86II::MRM_D9: case X86II::MRM_DA:
1595   case X86II::MRM_DB: case X86II::MRM_DC: case X86II::MRM_DD:
1596   case X86II::MRM_DE: case X86II::MRM_DF: case X86II::MRM_E0:
1597   case X86II::MRM_E1: case X86II::MRM_E2: case X86II::MRM_E3:
1598   case X86II::MRM_E4: case X86II::MRM_E5: case X86II::MRM_E6:
1599   case X86II::MRM_E7: case X86II::MRM_E8: case X86II::MRM_E9:
1600   case X86II::MRM_EA: case X86II::MRM_EB: case X86II::MRM_EC:
1601   case X86II::MRM_ED: case X86II::MRM_EE: case X86II::MRM_EF:
1602   case X86II::MRM_F0: case X86II::MRM_F1: case X86II::MRM_F2:
1603   case X86II::MRM_F3: case X86II::MRM_F4: case X86II::MRM_F5:
1604   case X86II::MRM_F6: case X86II::MRM_F7: case X86II::MRM_F8:
1605   case X86II::MRM_F9: case X86II::MRM_FA: case X86II::MRM_FB:
1606   case X86II::MRM_FC: case X86II::MRM_FD: case X86II::MRM_FE:
1607   case X86II::MRM_FF:
1608     EmitByte(BaseOpcode, CurByte, OS);
1609     EmitByte(0xC0 + Form - X86II::MRM_C0, CurByte, OS);
1610     break;
1611   }
1612 
1613   if (HasVEX_I8Reg) {
1614     // The last source register of a 4 operand instruction in AVX is encoded
1615     // in bits[7:4] of a immediate byte.
1616     assert(I8RegNum < 16 && "Register encoding out of range");
1617     I8RegNum <<= 4;
1618     if (CurOp != NumOps) {
1619       unsigned Val = MI.getOperand(CurOp++).getImm();
1620       assert(Val < 16 && "Immediate operand value out of range");
1621       I8RegNum |= Val;
1622     }
1623     EmitImmediate(MCOperand::createImm(I8RegNum), MI.getLoc(), 1, FK_Data_1,
1624                   CurByte, OS, Fixups);
1625   } else {
1626     // If there is a remaining operand, it must be a trailing immediate. Emit it
1627     // according to the right size for the instruction. Some instructions
1628     // (SSE4a extrq and insertq) have two trailing immediates.
1629     while (CurOp != NumOps && NumOps - CurOp <= 2) {
1630       EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1631                     X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1632                     CurByte, OS, Fixups);
1633     }
1634   }
1635 
1636   if ((TSFlags & X86II::OpMapMask) == X86II::ThreeDNow)
1637     EmitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS);
1638 
1639 #ifndef NDEBUG
1640   // FIXME: Verify.
1641   if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
1642     errs() << "Cannot encode all operands of: ";
1643     MI.dump();
1644     errs() << '\n';
1645     abort();
1646   }
1647 #endif
1648 }
1649 
1650 MCCodeEmitter *llvm::createX86MCCodeEmitter(const MCInstrInfo &MCII,
1651                                             const MCRegisterInfo &MRI,
1652                                             MCContext &Ctx) {
1653   return new X86MCCodeEmitter(MCII, Ctx);
1654 }
1655