xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp (revision ec4deee4e4f2aef1b97d9424f25d04e91fd7dc10)
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   X86MCCodeEmitter(const X86MCCodeEmitter &) = delete;
47   X86MCCodeEmitter &operator=(const X86MCCodeEmitter &) = delete;
48   ~X86MCCodeEmitter() override = default;
49 
50   void emitPrefix(const MCInst &MI, raw_ostream &OS,
51                   const MCSubtargetInfo &STI) const override;
52 
53   void encodeInstruction(const MCInst &MI, raw_ostream &OS,
54                          SmallVectorImpl<MCFixup> &Fixups,
55                          const MCSubtargetInfo &STI) const override;
56 
57 private:
58   unsigned getX86RegNum(const MCOperand &MO) const {
59     return Ctx.getRegisterInfo()->getEncodingValue(MO.getReg()) & 0x7;
60   }
61 
62   unsigned getX86RegEncoding(const MCInst &MI, unsigned OpNum) const {
63     return Ctx.getRegisterInfo()->getEncodingValue(
64         MI.getOperand(OpNum).getReg());
65   }
66 
67   /// \param MI a single low-level machine instruction.
68   /// \param OpNum the operand #.
69   /// \returns true if the OpNumth operand of MI  require a bit to be set in
70   /// REX prefix.
71   bool isREXExtendedReg(const MCInst &MI, unsigned OpNum) const {
72     return (getX86RegEncoding(MI, OpNum) >> 3) & 1;
73   }
74 
75   void emitByte(uint8_t C, unsigned &CurByte, raw_ostream &OS) const {
76     OS << (char)C;
77     ++CurByte;
78   }
79 
80   void emitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
81                     raw_ostream &OS) const {
82     // Output the constant in little endian byte order.
83     for (unsigned i = 0; i != Size; ++i) {
84       emitByte(Val & 255, CurByte, OS);
85       Val >>= 8;
86     }
87   }
88 
89   void emitImmediate(const MCOperand &Disp, SMLoc Loc, unsigned ImmSize,
90                      MCFixupKind FixupKind, unsigned &CurByte, raw_ostream &OS,
91                      SmallVectorImpl<MCFixup> &Fixups, int ImmOffset = 0) const;
92 
93   static uint8_t modRMByte(unsigned Mod, unsigned RegOpcode, unsigned RM) {
94     assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
95     return RM | (RegOpcode << 3) | (Mod << 6);
96   }
97 
98   void emitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
99                         unsigned &CurByte, raw_ostream &OS) const {
100     emitByte(modRMByte(3, RegOpcodeFld, getX86RegNum(ModRMReg)), CurByte, OS);
101   }
102 
103   void emitSIBByte(unsigned SS, unsigned Index, unsigned Base,
104                    unsigned &CurByte, raw_ostream &OS) const {
105     // SIB byte is in the same format as the modRMByte.
106     emitByte(modRMByte(SS, Index, Base), CurByte, OS);
107   }
108 
109   void emitMemModRMByte(const MCInst &MI, unsigned Op, unsigned RegOpcodeField,
110                         uint64_t TSFlags, bool Rex, unsigned &CurByte,
111                         raw_ostream &OS, SmallVectorImpl<MCFixup> &Fixups,
112                         const MCSubtargetInfo &STI) const;
113 
114   void emitPrefixImpl(uint64_t TSFlags, unsigned &CurOp, unsigned &CurByte,
115                   bool &Rex, const MCInst &MI, const MCInstrDesc &Desc,
116                   const MCSubtargetInfo &STI, raw_ostream &OS) const;
117 
118   void emitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
119                            const MCInst &MI, const MCInstrDesc &Desc,
120                            raw_ostream &OS) const;
121 
122   void emitSegmentOverridePrefix(unsigned &CurByte, unsigned SegOperand,
123                                  const MCInst &MI, raw_ostream &OS) const;
124 
125   bool emitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
126                         const MCInst &MI, const MCInstrDesc &Desc,
127                         const MCSubtargetInfo &STI, raw_ostream &OS) const;
128 
129   uint8_t determineREXPrefix(const MCInst &MI, uint64_t TSFlags, int MemOperand,
130                              const MCInstrDesc &Desc) const;
131 };
132 
133 } // end anonymous namespace
134 
135 /// \returns true if this signed displacement fits in a 8-bit sign-extended
136 /// field.
137 static bool isDisp8(int Value) { return Value == (int8_t)Value; }
138 
139 /// \returns true if this signed displacement fits in a 8-bit compressed
140 /// dispacement field.
141 static bool isCDisp8(uint64_t TSFlags, int Value, int &CValue) {
142   assert(((TSFlags & X86II::EncodingMask) == X86II::EVEX) &&
143          "Compressed 8-bit displacement is only valid for EVEX inst.");
144 
145   unsigned CD8_Scale =
146       (TSFlags & X86II::CD8_Scale_Mask) >> X86II::CD8_Scale_Shift;
147   if (CD8_Scale == 0) {
148     CValue = Value;
149     return isDisp8(Value);
150   }
151 
152   unsigned Mask = CD8_Scale - 1;
153   assert((CD8_Scale & Mask) == 0 && "Invalid memory object size.");
154   if (Value & Mask) // Unaligned offset
155     return false;
156   Value /= (int)CD8_Scale;
157   bool Ret = (Value == (int8_t)Value);
158 
159   if (Ret)
160     CValue = Value;
161   return Ret;
162 }
163 
164 /// \returns the appropriate fixup kind to use for an immediate in an
165 /// instruction with the specified TSFlags.
166 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
167   unsigned Size = X86II::getSizeOfImm(TSFlags);
168   bool isPCRel = X86II::isImmPCRel(TSFlags);
169 
170   if (X86II::isImmSigned(TSFlags)) {
171     switch (Size) {
172     default:
173       llvm_unreachable("Unsupported signed fixup size!");
174     case 4:
175       return MCFixupKind(X86::reloc_signed_4byte);
176     }
177   }
178   return MCFixup::getKindForSize(Size, isPCRel);
179 }
180 
181 /// \param Op operand # of the memory operand.
182 ///
183 /// \returns true if the specified instruction has a 16-bit memory operand.
184 static bool is16BitMemOperand(const MCInst &MI, unsigned Op,
185                               const MCSubtargetInfo &STI) {
186   const MCOperand &BaseReg = MI.getOperand(Op + X86::AddrBaseReg);
187   const MCOperand &IndexReg = MI.getOperand(Op + X86::AddrIndexReg);
188   const MCOperand &Disp = MI.getOperand(Op + X86::AddrDisp);
189 
190   if (STI.hasFeature(X86::Mode16Bit) && BaseReg.getReg() == 0 && Disp.isImm() &&
191       Disp.getImm() < 0x10000)
192     return true;
193   if ((BaseReg.getReg() != 0 &&
194        X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) ||
195       (IndexReg.getReg() != 0 &&
196        X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg())))
197     return true;
198   return false;
199 }
200 
201 /// \param Op operand # of the memory operand.
202 ///
203 /// \returns true if the specified instruction has a 32-bit memory operand.
204 static bool is32BitMemOperand(const MCInst &MI, unsigned Op) {
205   const MCOperand &BaseReg = MI.getOperand(Op + X86::AddrBaseReg);
206   const MCOperand &IndexReg = MI.getOperand(Op + X86::AddrIndexReg);
207 
208   if ((BaseReg.getReg() != 0 &&
209        X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) ||
210       (IndexReg.getReg() != 0 &&
211        X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg())))
212     return true;
213   if (BaseReg.getReg() == X86::EIP) {
214     assert(IndexReg.getReg() == 0 && "Invalid eip-based address.");
215     return true;
216   }
217   if (IndexReg.getReg() == X86::EIZ)
218     return true;
219   return false;
220 }
221 
222 /// \param Op operand # of the memory operand.
223 ///
224 /// \returns true if the specified instruction has a 64-bit memory operand.
225 #ifndef NDEBUG
226 static bool is64BitMemOperand(const MCInst &MI, unsigned Op) {
227   const MCOperand &BaseReg = MI.getOperand(Op + X86::AddrBaseReg);
228   const MCOperand &IndexReg = MI.getOperand(Op + X86::AddrIndexReg);
229 
230   if ((BaseReg.getReg() != 0 &&
231        X86MCRegisterClasses[X86::GR64RegClassID].contains(BaseReg.getReg())) ||
232       (IndexReg.getReg() != 0 &&
233        X86MCRegisterClasses[X86::GR64RegClassID].contains(IndexReg.getReg())))
234     return true;
235   return false;
236 }
237 #endif
238 
239 enum GlobalOffsetTableExprKind { GOT_None, GOT_Normal, GOT_SymDiff };
240 
241 /// Check if this expression starts with  _GLOBAL_OFFSET_TABLE_ and if it is
242 /// of the form _GLOBAL_OFFSET_TABLE_-symbol. This is needed to support PIC on
243 /// ELF i386 as _GLOBAL_OFFSET_TABLE_ is magical. We check only simple case that
244 /// are know to be used: _GLOBAL_OFFSET_TABLE_ by itself or at the start of a
245 /// binary expression.
246 static GlobalOffsetTableExprKind
247 startsWithGlobalOffsetTable(const MCExpr *Expr) {
248   const MCExpr *RHS = nullptr;
249   if (Expr->getKind() == MCExpr::Binary) {
250     const MCBinaryExpr *BE = static_cast<const MCBinaryExpr *>(Expr);
251     Expr = BE->getLHS();
252     RHS = BE->getRHS();
253   }
254 
255   if (Expr->getKind() != MCExpr::SymbolRef)
256     return GOT_None;
257 
258   const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr *>(Expr);
259   const MCSymbol &S = Ref->getSymbol();
260   if (S.getName() != "_GLOBAL_OFFSET_TABLE_")
261     return GOT_None;
262   if (RHS && RHS->getKind() == MCExpr::SymbolRef)
263     return GOT_SymDiff;
264   return GOT_Normal;
265 }
266 
267 static bool hasSecRelSymbolRef(const MCExpr *Expr) {
268   if (Expr->getKind() == MCExpr::SymbolRef) {
269     const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr *>(Expr);
270     return Ref->getKind() == MCSymbolRefExpr::VK_SECREL;
271   }
272   return false;
273 }
274 
275 static bool isPCRel32Branch(const MCInst &MI, const MCInstrInfo &MCII) {
276   unsigned Opcode = MI.getOpcode();
277   const MCInstrDesc &Desc = MCII.get(Opcode);
278   if ((Opcode != X86::CALL64pcrel32 && Opcode != X86::JMP_4) ||
279       getImmFixupKind(Desc.TSFlags) != FK_PCRel_4)
280     return false;
281 
282   unsigned CurOp = X86II::getOperandBias(Desc);
283   const MCOperand &Op = MI.getOperand(CurOp);
284   if (!Op.isExpr())
285     return false;
286 
287   const MCSymbolRefExpr *Ref = dyn_cast<MCSymbolRefExpr>(Op.getExpr());
288   return Ref && Ref->getKind() == MCSymbolRefExpr::VK_None;
289 }
290 
291 void X86MCCodeEmitter::emitImmediate(const MCOperand &DispOp, SMLoc Loc,
292                                      unsigned Size, MCFixupKind FixupKind,
293                                      unsigned &CurByte, raw_ostream &OS,
294                                      SmallVectorImpl<MCFixup> &Fixups,
295                                      int ImmOffset) const {
296   const MCExpr *Expr = nullptr;
297   if (DispOp.isImm()) {
298     // If this is a simple integer displacement that doesn't require a
299     // relocation, emit it now.
300     if (FixupKind != FK_PCRel_1 && FixupKind != FK_PCRel_2 &&
301         FixupKind != FK_PCRel_4) {
302       emitConstant(DispOp.getImm() + ImmOffset, Size, CurByte, OS);
303       return;
304     }
305     Expr = MCConstantExpr::create(DispOp.getImm(), Ctx);
306   } else {
307     Expr = DispOp.getExpr();
308   }
309 
310   // If we have an immoffset, add it to the expression.
311   if ((FixupKind == FK_Data_4 || FixupKind == FK_Data_8 ||
312        FixupKind == MCFixupKind(X86::reloc_signed_4byte))) {
313     GlobalOffsetTableExprKind Kind = startsWithGlobalOffsetTable(Expr);
314     if (Kind != GOT_None) {
315       assert(ImmOffset == 0);
316 
317       if (Size == 8) {
318         FixupKind = MCFixupKind(X86::reloc_global_offset_table8);
319       } else {
320         assert(Size == 4);
321         FixupKind = MCFixupKind(X86::reloc_global_offset_table);
322       }
323 
324       if (Kind == GOT_Normal)
325         ImmOffset = CurByte;
326     } else if (Expr->getKind() == MCExpr::SymbolRef) {
327       if (hasSecRelSymbolRef(Expr)) {
328         FixupKind = MCFixupKind(FK_SecRel_4);
329       }
330     } else if (Expr->getKind() == MCExpr::Binary) {
331       const MCBinaryExpr *Bin = static_cast<const MCBinaryExpr *>(Expr);
332       if (hasSecRelSymbolRef(Bin->getLHS()) ||
333           hasSecRelSymbolRef(Bin->getRHS())) {
334         FixupKind = MCFixupKind(FK_SecRel_4);
335       }
336     }
337   }
338 
339   // If the fixup is pc-relative, we need to bias the value to be relative to
340   // the start of the field, not the end of the field.
341   if (FixupKind == FK_PCRel_4 ||
342       FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
343       FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load) ||
344       FixupKind == MCFixupKind(X86::reloc_riprel_4byte_relax) ||
345       FixupKind == MCFixupKind(X86::reloc_riprel_4byte_relax_rex) ||
346       FixupKind == MCFixupKind(X86::reloc_branch_4byte_pcrel)) {
347     ImmOffset -= 4;
348     // If this is a pc-relative load off _GLOBAL_OFFSET_TABLE_:
349     // leaq _GLOBAL_OFFSET_TABLE_(%rip), %r15
350     // this needs to be a GOTPC32 relocation.
351     if (startsWithGlobalOffsetTable(Expr) != GOT_None)
352       FixupKind = MCFixupKind(X86::reloc_global_offset_table);
353   }
354   if (FixupKind == FK_PCRel_2)
355     ImmOffset -= 2;
356   if (FixupKind == FK_PCRel_1)
357     ImmOffset -= 1;
358 
359   if (ImmOffset)
360     Expr = MCBinaryExpr::createAdd(Expr, MCConstantExpr::create(ImmOffset, Ctx),
361                                    Ctx);
362 
363   // Emit a symbolic constant as a fixup and 4 zeros.
364   Fixups.push_back(MCFixup::create(CurByte, Expr, FixupKind, Loc));
365   emitConstant(0, Size, CurByte, OS);
366 }
367 
368 void X86MCCodeEmitter::emitMemModRMByte(const MCInst &MI, unsigned Op,
369                                         unsigned RegOpcodeField,
370                                         uint64_t TSFlags, bool Rex,
371                                         unsigned &CurByte, raw_ostream &OS,
372                                         SmallVectorImpl<MCFixup> &Fixups,
373                                         const MCSubtargetInfo &STI) const {
374   const MCOperand &Disp = MI.getOperand(Op + X86::AddrDisp);
375   const MCOperand &Base = MI.getOperand(Op + X86::AddrBaseReg);
376   const MCOperand &Scale = MI.getOperand(Op + X86::AddrScaleAmt);
377   const MCOperand &IndexReg = MI.getOperand(Op + X86::AddrIndexReg);
378   unsigned BaseReg = Base.getReg();
379   bool HasEVEX = (TSFlags & X86II::EncodingMask) == X86II::EVEX;
380 
381   // Handle %rip relative addressing.
382   if (BaseReg == X86::RIP ||
383       BaseReg == X86::EIP) { // [disp32+rIP] in X86-64 mode
384     assert(STI.hasFeature(X86::Mode64Bit) &&
385            "Rip-relative addressing requires 64-bit mode");
386     assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
387     emitByte(modRMByte(0, RegOpcodeField, 5), CurByte, OS);
388 
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 = [=]() {
394       switch (Opcode) {
395       default:
396         return X86::reloc_riprel_4byte;
397       case X86::MOV64rm:
398         assert(Rex);
399         return X86::reloc_riprel_4byte_movq_load;
400       case X86::CALL64m:
401       case X86::JMP64m:
402       case X86::TAILJMPm64:
403       case X86::TEST64mr:
404       case X86::ADC64rm:
405       case X86::ADD64rm:
406       case X86::AND64rm:
407       case X86::CMP64rm:
408       case X86::OR64rm:
409       case X86::SBB64rm:
410       case X86::SUB64rm:
411       case X86::XOR64rm:
412         return Rex ? X86::reloc_riprel_4byte_relax_rex
413                    : X86::reloc_riprel_4byte_relax;
414       }
415     }();
416 
417     // rip-relative addressing is actually relative to the *next* instruction.
418     // Since an immediate can follow the mod/rm byte for an instruction, this
419     // means that we need to bias the displacement field of the instruction with
420     // the size of the immediate field. If we have this case, add it into the
421     // expression to emit.
422     // Note: rip-relative addressing using immediate displacement values should
423     // not be adjusted, assuming it was the user's intent.
424     int ImmSize = !Disp.isImm() && X86II::hasImm(TSFlags)
425                       ? X86II::getSizeOfImm(TSFlags)
426                       : 0;
427 
428     emitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind), CurByte, OS,
429                   Fixups, -ImmSize);
430     return;
431   }
432 
433   unsigned BaseRegNo = BaseReg ? getX86RegNum(Base) : -1U;
434 
435   // 16-bit addressing forms of the ModR/M byte have a different encoding for
436   // the R/M field and are far more limited in which registers can be used.
437   if (is16BitMemOperand(MI, Op, STI)) {
438     if (BaseReg) {
439       // For 32-bit addressing, the row and column values in Table 2-2 are
440       // basically the same. It's AX/CX/DX/BX/SP/BP/SI/DI in that order, with
441       // some special cases. And getX86RegNum reflects that numbering.
442       // For 16-bit addressing it's more fun, as shown in the SDM Vol 2A,
443       // Table 2-1 "16-Bit Addressing Forms with the ModR/M byte". We can only
444       // use SI/DI/BP/BX, which have "row" values 4-7 in no particular order,
445       // while values 0-3 indicate the allowed combinations (base+index) of
446       // those: 0 for BX+SI, 1 for BX+DI, 2 for BP+SI, 3 for BP+DI.
447       //
448       // R16Table[] is a lookup from the normal RegNo, to the row values from
449       // Table 2-1 for 16-bit addressing modes. Where zero means disallowed.
450       static const unsigned R16Table[] = {0, 0, 0, 7, 0, 6, 4, 5};
451       unsigned RMfield = R16Table[BaseRegNo];
452 
453       assert(RMfield && "invalid 16-bit base register");
454 
455       if (IndexReg.getReg()) {
456         unsigned IndexReg16 = R16Table[getX86RegNum(IndexReg)];
457 
458         assert(IndexReg16 && "invalid 16-bit index register");
459         // We must have one of SI/DI (4,5), and one of BP/BX (6,7).
460         assert(((IndexReg16 ^ RMfield) & 2) &&
461                "invalid 16-bit base/index register combination");
462         assert(Scale.getImm() == 1 &&
463                "invalid scale for 16-bit memory reference");
464 
465         // Allow base/index to appear in either order (although GAS doesn't).
466         if (IndexReg16 & 2)
467           RMfield = (RMfield & 1) | ((7 - IndexReg16) << 1);
468         else
469           RMfield = (IndexReg16 & 1) | ((7 - RMfield) << 1);
470       }
471 
472       if (Disp.isImm() && isDisp8(Disp.getImm())) {
473         if (Disp.getImm() == 0 && RMfield != 6) {
474           // There is no displacement; just the register.
475           emitByte(modRMByte(0, RegOpcodeField, RMfield), CurByte, OS);
476           return;
477         }
478         // Use the [REG]+disp8 form, including for [BP] which cannot be encoded.
479         emitByte(modRMByte(1, RegOpcodeField, RMfield), CurByte, OS);
480         emitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
481         return;
482       }
483       // This is the [REG]+disp16 case.
484       emitByte(modRMByte(2, RegOpcodeField, RMfield), CurByte, OS);
485     } else {
486       // There is no BaseReg; this is the plain [disp16] case.
487       emitByte(modRMByte(0, RegOpcodeField, 6), CurByte, OS);
488     }
489 
490     // Emit 16-bit displacement for plain disp16 or [REG]+disp16 cases.
491     emitImmediate(Disp, MI.getLoc(), 2, FK_Data_2, CurByte, OS, Fixups);
492     return;
493   }
494 
495   // Determine whether a SIB byte is needed.
496   // If no BaseReg, issue a RIP relative instruction only if the MCE can
497   // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
498   // 2-7) and absolute references.
499 
500   if ( // The SIB byte must be used if there is an index register.
501       IndexReg.getReg() == 0 &&
502       // The SIB byte must be used if the base is ESP/RSP/R12, all of which
503       // encode to an R/M value of 4, which indicates that a SIB byte is
504       // present.
505       BaseRegNo != N86::ESP &&
506       // If there is no base register and we're in 64-bit mode, we need a SIB
507       // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
508       (!STI.hasFeature(X86::Mode64Bit) || BaseReg != 0)) {
509 
510     if (BaseReg == 0) { // [disp32]     in X86-32 mode
511       emitByte(modRMByte(0, RegOpcodeField, 5), CurByte, OS);
512       emitImmediate(Disp, MI.getLoc(), 4, FK_Data_4, CurByte, OS, Fixups);
513       return;
514     }
515 
516     // If the base is not EBP/ESP and there is no displacement, use simple
517     // indirect register encoding, this handles addresses like [EAX].  The
518     // encoding for [EBP] with no displacement means [disp32] so we handle it
519     // by emitting a displacement of 0 below.
520     if (BaseRegNo != N86::EBP) {
521       if (Disp.isImm() && Disp.getImm() == 0) {
522         emitByte(modRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
523         return;
524       }
525 
526       // If the displacement is @tlscall, treat it as a zero.
527       if (Disp.isExpr()) {
528         auto *Sym = dyn_cast<MCSymbolRefExpr>(Disp.getExpr());
529         if (Sym && Sym->getKind() == MCSymbolRefExpr::VK_TLSCALL) {
530           // This is exclusively used by call *a@tlscall(base). The relocation
531           // (R_386_TLSCALL or R_X86_64_TLSCALL) applies to the beginning.
532           Fixups.push_back(MCFixup::create(0, Sym, FK_NONE, MI.getLoc()));
533           emitByte(modRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
534           return;
535         }
536       }
537     }
538 
539     // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
540     if (Disp.isImm()) {
541       if (!HasEVEX && isDisp8(Disp.getImm())) {
542         emitByte(modRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
543         emitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
544         return;
545       }
546       // Try EVEX compressed 8-bit displacement first; if failed, fall back to
547       // 32-bit displacement.
548       int CDisp8 = 0;
549       if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
550         emitByte(modRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
551         emitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups,
552                       CDisp8 - Disp.getImm());
553         return;
554       }
555     }
556 
557     // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
558     emitByte(modRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
559     unsigned Opcode = MI.getOpcode();
560     unsigned FixupKind = Opcode == X86::MOV32rm ? X86::reloc_signed_4byte_relax
561                                                 : X86::reloc_signed_4byte;
562     emitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind), CurByte, OS,
563                   Fixups);
564     return;
565   }
566 
567   // We need a SIB byte, so start by outputting the ModR/M byte first
568   assert(IndexReg.getReg() != X86::ESP && IndexReg.getReg() != X86::RSP &&
569          "Cannot use ESP as index reg!");
570 
571   bool ForceDisp32 = false;
572   bool ForceDisp8 = false;
573   int CDisp8 = 0;
574   int ImmOffset = 0;
575   if (BaseReg == 0) {
576     // If there is no base register, we emit the special case SIB byte with
577     // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
578     emitByte(modRMByte(0, RegOpcodeField, 4), CurByte, OS);
579     ForceDisp32 = true;
580   } else if (!Disp.isImm()) {
581     // Emit the normal disp32 encoding.
582     emitByte(modRMByte(2, RegOpcodeField, 4), CurByte, OS);
583     ForceDisp32 = true;
584   } else if (Disp.getImm() == 0 &&
585              // Base reg can't be anything that ends up with '5' as the base
586              // reg, it is the magic [*] nomenclature that indicates no base.
587              BaseRegNo != N86::EBP) {
588     // Emit no displacement ModR/M byte
589     emitByte(modRMByte(0, RegOpcodeField, 4), CurByte, OS);
590   } else if (!HasEVEX && isDisp8(Disp.getImm())) {
591     // Emit the disp8 encoding.
592     emitByte(modRMByte(1, RegOpcodeField, 4), CurByte, OS);
593     ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
594   } else if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
595     // Emit the disp8 encoding.
596     emitByte(modRMByte(1, RegOpcodeField, 4), CurByte, OS);
597     ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
598     ImmOffset = CDisp8 - Disp.getImm();
599   } else {
600     // Emit the normal disp32 encoding.
601     emitByte(modRMByte(2, RegOpcodeField, 4), CurByte, OS);
602   }
603 
604   // Calculate what the SS field value should be...
605   static const unsigned SSTable[] = {~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3};
606   unsigned SS = SSTable[Scale.getImm()];
607 
608   if (BaseReg == 0) {
609     // Handle the SIB byte for the case where there is no base, see Intel
610     // Manual 2A, table 2-7. The displacement has already been output.
611     unsigned IndexRegNo;
612     if (IndexReg.getReg())
613       IndexRegNo = getX86RegNum(IndexReg);
614     else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
615       IndexRegNo = 4;
616     emitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
617   } else {
618     unsigned IndexRegNo;
619     if (IndexReg.getReg())
620       IndexRegNo = getX86RegNum(IndexReg);
621     else
622       IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
623     emitSIBByte(SS, IndexRegNo, getX86RegNum(Base), CurByte, OS);
624   }
625 
626   // Do we need to output a displacement?
627   if (ForceDisp8)
628     emitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups,
629                   ImmOffset);
630   else if (ForceDisp32 || Disp.getImm() != 0)
631     emitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(X86::reloc_signed_4byte),
632                   CurByte, OS, Fixups);
633 }
634 
635 void X86MCCodeEmitter::emitPrefixImpl(uint64_t TSFlags, unsigned &CurOp,
636                                   unsigned &CurByte, bool &Rex,
637                                   const MCInst &MI, const MCInstrDesc &Desc,
638                                   const MCSubtargetInfo &STI,
639                                   raw_ostream &OS) const {
640   // Determine where the memory operand starts, if present.
641   int MemoryOperand = X86II::getMemoryOperandNo(TSFlags);
642   if (MemoryOperand != -1)
643     MemoryOperand += CurOp;
644 
645   // Emit segment override opcode prefix as needed.
646   if (MemoryOperand >= 0)
647     emitSegmentOverridePrefix(CurByte, MemoryOperand + X86::AddrSegmentReg, MI,
648                               OS);
649 
650   // Emit the repeat opcode prefix as needed.
651   unsigned Flags = MI.getFlags();
652   if (TSFlags & X86II::REP || Flags & X86::IP_HAS_REPEAT)
653     emitByte(0xF3, CurByte, OS);
654   if (Flags & X86::IP_HAS_REPEAT_NE)
655     emitByte(0xF2, CurByte, OS);
656 
657   // Emit the address size opcode prefix as needed.
658   bool need_address_override;
659   uint64_t AdSize = TSFlags & X86II::AdSizeMask;
660   if ((STI.hasFeature(X86::Mode16Bit) && AdSize == X86II::AdSize32) ||
661       (STI.hasFeature(X86::Mode32Bit) && AdSize == X86II::AdSize16) ||
662       (STI.hasFeature(X86::Mode64Bit) && AdSize == X86II::AdSize32)) {
663     need_address_override = true;
664   } else if (MemoryOperand < 0) {
665     need_address_override = false;
666   } else if (STI.hasFeature(X86::Mode64Bit)) {
667     assert(!is16BitMemOperand(MI, MemoryOperand, STI));
668     need_address_override = is32BitMemOperand(MI, MemoryOperand);
669   } else if (STI.hasFeature(X86::Mode32Bit)) {
670     assert(!is64BitMemOperand(MI, MemoryOperand));
671     need_address_override = is16BitMemOperand(MI, MemoryOperand, STI);
672   } else {
673     assert(STI.hasFeature(X86::Mode16Bit));
674     assert(!is64BitMemOperand(MI, MemoryOperand));
675     need_address_override = !is16BitMemOperand(MI, MemoryOperand, STI);
676   }
677 
678   if (need_address_override)
679     emitByte(0x67, CurByte, OS);
680 
681   // Encoding type for this instruction.
682   uint64_t Encoding = TSFlags & X86II::EncodingMask;
683   if (Encoding == 0)
684     Rex = emitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, STI, OS);
685   else
686     emitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
687 
688   uint64_t Form = TSFlags & X86II::FormMask;
689   switch (Form) {
690   default:
691     break;
692   case X86II::RawFrmDstSrc: {
693     unsigned siReg = MI.getOperand(1).getReg();
694     assert(((siReg == X86::SI && MI.getOperand(0).getReg() == X86::DI) ||
695             (siReg == X86::ESI && MI.getOperand(0).getReg() == X86::EDI) ||
696             (siReg == X86::RSI && MI.getOperand(0).getReg() == X86::RDI)) &&
697            "SI and DI register sizes do not match");
698     // Emit segment override opcode prefix as needed (not for %ds).
699     if (MI.getOperand(2).getReg() != X86::DS)
700       emitSegmentOverridePrefix(CurByte, 2, MI, OS);
701     // Emit AdSize prefix as needed.
702     if ((!STI.hasFeature(X86::Mode32Bit) && siReg == X86::ESI) ||
703         (STI.hasFeature(X86::Mode32Bit) && siReg == X86::SI))
704       emitByte(0x67, CurByte, OS);
705     CurOp += 3; // Consume operands.
706     break;
707   }
708   case X86II::RawFrmSrc: {
709     unsigned siReg = MI.getOperand(0).getReg();
710     // Emit segment override opcode prefix as needed (not for %ds).
711     if (MI.getOperand(1).getReg() != X86::DS)
712       emitSegmentOverridePrefix(CurByte, 1, MI, OS);
713     // Emit AdSize prefix as needed.
714     if ((!STI.hasFeature(X86::Mode32Bit) && siReg == X86::ESI) ||
715         (STI.hasFeature(X86::Mode32Bit) && siReg == X86::SI))
716       emitByte(0x67, CurByte, OS);
717     CurOp += 2; // Consume operands.
718     break;
719   }
720   case X86II::RawFrmDst: {
721     unsigned siReg = MI.getOperand(0).getReg();
722     // Emit AdSize prefix as needed.
723     if ((!STI.hasFeature(X86::Mode32Bit) && siReg == X86::EDI) ||
724         (STI.hasFeature(X86::Mode32Bit) && siReg == X86::DI))
725       emitByte(0x67, CurByte, OS);
726     ++CurOp; // Consume operand.
727     break;
728   }
729   case X86II::RawFrmMemOffs: {
730     // Emit segment override opcode prefix as needed.
731     emitSegmentOverridePrefix(CurByte, 1, MI, OS);
732     break;
733   }
734   }
735 }
736 
737 /// emitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
738 /// called VEX.
739 void X86MCCodeEmitter::emitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
740                                            int MemOperand, const MCInst &MI,
741                                            const MCInstrDesc &Desc,
742                                            raw_ostream &OS) const {
743   assert(!(TSFlags & X86II::LOCK) && "Can't have LOCK VEX.");
744 
745   uint64_t Encoding = TSFlags & X86II::EncodingMask;
746   bool HasEVEX_K = TSFlags & X86II::EVEX_K;
747   bool HasVEX_4V = TSFlags & X86II::VEX_4V;
748   bool HasEVEX_RC = TSFlags & X86II::EVEX_RC;
749 
750   // VEX_R: opcode externsion equivalent to REX.R in
751   // 1's complement (inverted) form
752   //
753   //  1: Same as REX_R=0 (must be 1 in 32-bit mode)
754   //  0: Same as REX_R=1 (64 bit mode only)
755   //
756   uint8_t VEX_R = 0x1;
757   uint8_t EVEX_R2 = 0x1;
758 
759   // VEX_X: equivalent to REX.X, only used when a
760   // register is used for index in SIB Byte.
761   //
762   //  1: Same as REX.X=0 (must be 1 in 32-bit mode)
763   //  0: Same as REX.X=1 (64-bit mode only)
764   uint8_t VEX_X = 0x1;
765 
766   // VEX_B:
767   //
768   //  1: Same as REX_B=0 (ignored in 32-bit mode)
769   //  0: Same as REX_B=1 (64 bit mode only)
770   //
771   uint8_t VEX_B = 0x1;
772 
773   // VEX_W: opcode specific (use like REX.W, or used for
774   // opcode extension, or ignored, depending on the opcode byte)
775   uint8_t VEX_W = (TSFlags & X86II::VEX_W) ? 1 : 0;
776 
777   // VEX_5M (VEX m-mmmmm field):
778   //
779   //  0b00000: Reserved for future use
780   //  0b00001: implied 0F leading opcode
781   //  0b00010: implied 0F 38 leading opcode bytes
782   //  0b00011: implied 0F 3A leading opcode bytes
783   //  0b00100-0b11111: Reserved for future use
784   //  0b01000: XOP map select - 08h instructions with imm byte
785   //  0b01001: XOP map select - 09h instructions with no imm byte
786   //  0b01010: XOP map select - 0Ah instructions with imm dword
787   uint8_t VEX_5M;
788   switch (TSFlags & X86II::OpMapMask) {
789   default:
790     llvm_unreachable("Invalid prefix!");
791   case X86II::TB:
792     VEX_5M = 0x1;
793     break; // 0F
794   case X86II::T8:
795     VEX_5M = 0x2;
796     break; // 0F 38
797   case X86II::TA:
798     VEX_5M = 0x3;
799     break; // 0F 3A
800   case X86II::XOP8:
801     VEX_5M = 0x8;
802     break;
803   case X86II::XOP9:
804     VEX_5M = 0x9;
805     break;
806   case X86II::XOPA:
807     VEX_5M = 0xA;
808     break;
809   }
810 
811   // VEX_4V (VEX vvvv field): a register specifier
812   // (in 1's complement form) or 1111 if unused.
813   uint8_t VEX_4V = 0xf;
814   uint8_t EVEX_V2 = 0x1;
815 
816   // EVEX_L2/VEX_L (Vector Length):
817   //
818   // L2 L
819   //  0 0: scalar or 128-bit vector
820   //  0 1: 256-bit vector
821   //  1 0: 512-bit vector
822   //
823   uint8_t VEX_L = (TSFlags & X86II::VEX_L) ? 1 : 0;
824   uint8_t EVEX_L2 = (TSFlags & X86II::EVEX_L2) ? 1 : 0;
825 
826   // VEX_PP: opcode extension providing equivalent
827   // functionality of a SIMD prefix
828   //
829   //  0b00: None
830   //  0b01: 66
831   //  0b10: F3
832   //  0b11: F2
833   //
834   uint8_t VEX_PP = 0;
835   switch (TSFlags & X86II::OpPrefixMask) {
836   case X86II::PD:
837     VEX_PP = 0x1;
838     break; // 66
839   case X86II::XS:
840     VEX_PP = 0x2;
841     break; // F3
842   case X86II::XD:
843     VEX_PP = 0x3;
844     break; // F2
845   }
846 
847   // EVEX_U
848   uint8_t EVEX_U = 1; // Always '1' so far
849 
850   // EVEX_z
851   uint8_t EVEX_z = (HasEVEX_K && (TSFlags & X86II::EVEX_Z)) ? 1 : 0;
852 
853   // EVEX_b
854   uint8_t EVEX_b = (TSFlags & X86II::EVEX_B) ? 1 : 0;
855 
856   // EVEX_rc
857   uint8_t EVEX_rc = 0;
858 
859   // EVEX_aaa
860   uint8_t EVEX_aaa = 0;
861 
862   bool EncodeRC = false;
863 
864   // Classify VEX_B, VEX_4V, VEX_R, VEX_X
865   unsigned NumOps = Desc.getNumOperands();
866   unsigned CurOp = X86II::getOperandBias(Desc);
867 
868   switch (TSFlags & X86II::FormMask) {
869   default:
870     llvm_unreachable("Unexpected form in emitVEXOpcodePrefix!");
871   case X86II::RawFrm:
872     break;
873   case X86II::MRMDestMem: {
874     // MRMDestMem instructions forms:
875     //  MemAddr, src1(ModR/M)
876     //  MemAddr, src1(VEX_4V), src2(ModR/M)
877     //  MemAddr, src1(ModR/M), imm8
878     //
879     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
880     VEX_B = ~(BaseRegEnc >> 3) & 1;
881     unsigned IndexRegEnc =
882         getX86RegEncoding(MI, MemOperand + X86::AddrIndexReg);
883     VEX_X = ~(IndexRegEnc >> 3) & 1;
884     if (!HasVEX_4V) // Only needed with VSIB which don't use VVVV.
885       EVEX_V2 = ~(IndexRegEnc >> 4) & 1;
886 
887     CurOp += X86::AddrNumOperands;
888 
889     if (HasEVEX_K)
890       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
891 
892     if (HasVEX_4V) {
893       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
894       VEX_4V = ~VRegEnc & 0xf;
895       EVEX_V2 = ~(VRegEnc >> 4) & 1;
896     }
897 
898     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
899     VEX_R = ~(RegEnc >> 3) & 1;
900     EVEX_R2 = ~(RegEnc >> 4) & 1;
901     break;
902   }
903   case X86II::MRMSrcMem: {
904     // MRMSrcMem instructions forms:
905     //  src1(ModR/M), MemAddr
906     //  src1(ModR/M), src2(VEX_4V), MemAddr
907     //  src1(ModR/M), MemAddr, imm8
908     //  src1(ModR/M), MemAddr, src2(Imm[7:4])
909     //
910     //  FMA4:
911     //  dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(Imm[7:4])
912     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
913     VEX_R = ~(RegEnc >> 3) & 1;
914     EVEX_R2 = ~(RegEnc >> 4) & 1;
915 
916     if (HasEVEX_K)
917       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
918 
919     if (HasVEX_4V) {
920       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
921       VEX_4V = ~VRegEnc & 0xf;
922       EVEX_V2 = ~(VRegEnc >> 4) & 1;
923     }
924 
925     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
926     VEX_B = ~(BaseRegEnc >> 3) & 1;
927     unsigned IndexRegEnc =
928         getX86RegEncoding(MI, MemOperand + X86::AddrIndexReg);
929     VEX_X = ~(IndexRegEnc >> 3) & 1;
930     if (!HasVEX_4V) // Only needed with VSIB which don't use VVVV.
931       EVEX_V2 = ~(IndexRegEnc >> 4) & 1;
932 
933     break;
934   }
935   case X86II::MRMSrcMem4VOp3: {
936     // Instruction format for 4VOp3:
937     //   src1(ModR/M), MemAddr, src3(VEX_4V)
938     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
939     VEX_R = ~(RegEnc >> 3) & 1;
940 
941     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
942     VEX_B = ~(BaseRegEnc >> 3) & 1;
943     unsigned IndexRegEnc =
944         getX86RegEncoding(MI, MemOperand + X86::AddrIndexReg);
945     VEX_X = ~(IndexRegEnc >> 3) & 1;
946 
947     VEX_4V = ~getX86RegEncoding(MI, CurOp + X86::AddrNumOperands) & 0xf;
948     break;
949   }
950   case X86II::MRMSrcMemOp4: {
951     //  dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
952     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
953     VEX_R = ~(RegEnc >> 3) & 1;
954 
955     unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
956     VEX_4V = ~VRegEnc & 0xf;
957 
958     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
959     VEX_B = ~(BaseRegEnc >> 3) & 1;
960     unsigned IndexRegEnc =
961         getX86RegEncoding(MI, MemOperand + X86::AddrIndexReg);
962     VEX_X = ~(IndexRegEnc >> 3) & 1;
963     break;
964   }
965   case X86II::MRM0m:
966   case X86II::MRM1m:
967   case X86II::MRM2m:
968   case X86II::MRM3m:
969   case X86II::MRM4m:
970   case X86II::MRM5m:
971   case X86II::MRM6m:
972   case X86II::MRM7m: {
973     // MRM[0-9]m instructions forms:
974     //  MemAddr
975     //  src1(VEX_4V), MemAddr
976     if (HasVEX_4V) {
977       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
978       VEX_4V = ~VRegEnc & 0xf;
979       EVEX_V2 = ~(VRegEnc >> 4) & 1;
980     }
981 
982     if (HasEVEX_K)
983       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
984 
985     unsigned BaseRegEnc = getX86RegEncoding(MI, MemOperand + X86::AddrBaseReg);
986     VEX_B = ~(BaseRegEnc >> 3) & 1;
987     unsigned IndexRegEnc =
988         getX86RegEncoding(MI, MemOperand + X86::AddrIndexReg);
989     VEX_X = ~(IndexRegEnc >> 3) & 1;
990     if (!HasVEX_4V) // Only needed with VSIB which don't use VVVV.
991       EVEX_V2 = ~(IndexRegEnc >> 4) & 1;
992 
993     break;
994   }
995   case X86II::MRMSrcReg: {
996     // MRMSrcReg instructions forms:
997     //  dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(Imm[7:4])
998     //  dst(ModR/M), src1(ModR/M)
999     //  dst(ModR/M), src1(ModR/M), imm8
1000     //
1001     //  FMA4:
1002     //  dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
1003     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
1004     VEX_R = ~(RegEnc >> 3) & 1;
1005     EVEX_R2 = ~(RegEnc >> 4) & 1;
1006 
1007     if (HasEVEX_K)
1008       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
1009 
1010     if (HasVEX_4V) {
1011       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
1012       VEX_4V = ~VRegEnc & 0xf;
1013       EVEX_V2 = ~(VRegEnc >> 4) & 1;
1014     }
1015 
1016     RegEnc = getX86RegEncoding(MI, CurOp++);
1017     VEX_B = ~(RegEnc >> 3) & 1;
1018     VEX_X = ~(RegEnc >> 4) & 1;
1019 
1020     if (EVEX_b) {
1021       if (HasEVEX_RC) {
1022         unsigned RcOperand = NumOps - 1;
1023         assert(RcOperand >= CurOp);
1024         EVEX_rc = MI.getOperand(RcOperand).getImm();
1025         assert(EVEX_rc <= 3 && "Invalid rounding control!");
1026       }
1027       EncodeRC = true;
1028     }
1029     break;
1030   }
1031   case X86II::MRMSrcReg4VOp3: {
1032     // Instruction format for 4VOp3:
1033     //   src1(ModR/M), src2(ModR/M), src3(VEX_4V)
1034     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
1035     VEX_R = ~(RegEnc >> 3) & 1;
1036 
1037     RegEnc = getX86RegEncoding(MI, CurOp++);
1038     VEX_B = ~(RegEnc >> 3) & 1;
1039 
1040     VEX_4V = ~getX86RegEncoding(MI, CurOp++) & 0xf;
1041     break;
1042   }
1043   case X86II::MRMSrcRegOp4: {
1044     //  dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
1045     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
1046     VEX_R = ~(RegEnc >> 3) & 1;
1047 
1048     unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
1049     VEX_4V = ~VRegEnc & 0xf;
1050 
1051     // Skip second register source (encoded in Imm[7:4])
1052     ++CurOp;
1053 
1054     RegEnc = getX86RegEncoding(MI, CurOp++);
1055     VEX_B = ~(RegEnc >> 3) & 1;
1056     VEX_X = ~(RegEnc >> 4) & 1;
1057     break;
1058   }
1059   case X86II::MRMDestReg: {
1060     // MRMDestReg instructions forms:
1061     //  dst(ModR/M), src(ModR/M)
1062     //  dst(ModR/M), src(ModR/M), imm8
1063     //  dst(ModR/M), src1(VEX_4V), src2(ModR/M)
1064     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
1065     VEX_B = ~(RegEnc >> 3) & 1;
1066     VEX_X = ~(RegEnc >> 4) & 1;
1067 
1068     if (HasEVEX_K)
1069       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
1070 
1071     if (HasVEX_4V) {
1072       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
1073       VEX_4V = ~VRegEnc & 0xf;
1074       EVEX_V2 = ~(VRegEnc >> 4) & 1;
1075     }
1076 
1077     RegEnc = getX86RegEncoding(MI, CurOp++);
1078     VEX_R = ~(RegEnc >> 3) & 1;
1079     EVEX_R2 = ~(RegEnc >> 4) & 1;
1080     if (EVEX_b)
1081       EncodeRC = true;
1082     break;
1083   }
1084   case X86II::MRM0r:
1085   case X86II::MRM1r:
1086   case X86II::MRM2r:
1087   case X86II::MRM3r:
1088   case X86II::MRM4r:
1089   case X86II::MRM5r:
1090   case X86II::MRM6r:
1091   case X86II::MRM7r: {
1092     // MRM0r-MRM7r instructions forms:
1093     //  dst(VEX_4V), src(ModR/M), imm8
1094     if (HasVEX_4V) {
1095       unsigned VRegEnc = getX86RegEncoding(MI, CurOp++);
1096       VEX_4V = ~VRegEnc & 0xf;
1097       EVEX_V2 = ~(VRegEnc >> 4) & 1;
1098     }
1099     if (HasEVEX_K)
1100       EVEX_aaa = getX86RegEncoding(MI, CurOp++);
1101 
1102     unsigned RegEnc = getX86RegEncoding(MI, CurOp++);
1103     VEX_B = ~(RegEnc >> 3) & 1;
1104     VEX_X = ~(RegEnc >> 4) & 1;
1105     break;
1106   }
1107   }
1108 
1109   if (Encoding == X86II::VEX || Encoding == X86II::XOP) {
1110     // VEX opcode prefix can have 2 or 3 bytes
1111     //
1112     //  3 bytes:
1113     //    +-----+ +--------------+ +-------------------+
1114     //    | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
1115     //    +-----+ +--------------+ +-------------------+
1116     //  2 bytes:
1117     //    +-----+ +-------------------+
1118     //    | C5h | | R | vvvv | L | pp |
1119     //    +-----+ +-------------------+
1120     //
1121     //  XOP uses a similar prefix:
1122     //    +-----+ +--------------+ +-------------------+
1123     //    | 8Fh | | RXB | m-mmmm | | W | vvvv | L | pp |
1124     //    +-----+ +--------------+ +-------------------+
1125     uint8_t LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
1126 
1127     // Can we use the 2 byte VEX prefix?
1128     if (!(MI.getFlags() & X86::IP_USE_VEX3) && Encoding == X86II::VEX &&
1129         VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) {
1130       emitByte(0xC5, CurByte, OS);
1131       emitByte(LastByte | (VEX_R << 7), CurByte, OS);
1132       return;
1133     }
1134 
1135     // 3 byte VEX prefix
1136     emitByte(Encoding == X86II::XOP ? 0x8F : 0xC4, CurByte, OS);
1137     emitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
1138     emitByte(LastByte | (VEX_W << 7), CurByte, OS);
1139   } else {
1140     assert(Encoding == X86II::EVEX && "unknown encoding!");
1141     // EVEX opcode prefix can have 4 bytes
1142     //
1143     // +-----+ +--------------+ +-------------------+ +------------------------+
1144     // | 62h | | RXBR' | 00mm | | W | vvvv | U | pp | | z | L'L | b | v' | aaa |
1145     // +-----+ +--------------+ +-------------------+ +------------------------+
1146     assert((VEX_5M & 0x3) == VEX_5M &&
1147            "More than 2 significant bits in VEX.m-mmmm fields for EVEX!");
1148 
1149     emitByte(0x62, CurByte, OS);
1150     emitByte((VEX_R << 7) | (VEX_X << 6) | (VEX_B << 5) | (EVEX_R2 << 4) |
1151                  VEX_5M,
1152              CurByte, OS);
1153     emitByte((VEX_W << 7) | (VEX_4V << 3) | (EVEX_U << 2) | VEX_PP, CurByte,
1154              OS);
1155     if (EncodeRC)
1156       emitByte((EVEX_z << 7) | (EVEX_rc << 5) | (EVEX_b << 4) | (EVEX_V2 << 3) |
1157                    EVEX_aaa,
1158                CurByte, OS);
1159     else
1160       emitByte((EVEX_z << 7) | (EVEX_L2 << 6) | (VEX_L << 5) | (EVEX_b << 4) |
1161                    (EVEX_V2 << 3) | EVEX_aaa,
1162                CurByte, OS);
1163   }
1164 }
1165 
1166 /// Determine if the MCInst has to be encoded with a X86-64 REX prefix which
1167 /// specifies 1) 64-bit instructions, 2) non-default operand size, and 3) use
1168 /// of X86-64 extended registers.
1169 uint8_t X86MCCodeEmitter::determineREXPrefix(const MCInst &MI, uint64_t TSFlags,
1170                                              int MemOperand,
1171                                              const MCInstrDesc &Desc) const {
1172   uint8_t REX = 0;
1173   bool UsesHighByteReg = false;
1174 
1175   if (TSFlags & X86II::REX_W)
1176     REX |= 1 << 3; // set REX.W
1177 
1178   if (MI.getNumOperands() == 0)
1179     return REX;
1180 
1181   unsigned NumOps = MI.getNumOperands();
1182   unsigned CurOp = X86II::getOperandBias(Desc);
1183 
1184   // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
1185   for (unsigned i = CurOp; i != NumOps; ++i) {
1186     const MCOperand &MO = MI.getOperand(i);
1187     if (!MO.isReg())
1188       continue;
1189     unsigned Reg = MO.getReg();
1190     if (Reg == X86::AH || Reg == X86::BH || Reg == X86::CH || Reg == X86::DH)
1191       UsesHighByteReg = true;
1192     if (X86II::isX86_64NonExtLowByteReg(Reg))
1193       // FIXME: The caller of determineREXPrefix slaps this prefix onto anything
1194       // that returns non-zero.
1195       REX |= 0x40; // REX fixed encoding prefix
1196   }
1197 
1198   switch (TSFlags & X86II::FormMask) {
1199   case X86II::AddRegFrm:
1200     REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1201     break;
1202   case X86II::MRMSrcReg:
1203   case X86II::MRMSrcRegCC:
1204     REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1205     REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1206     break;
1207   case X86II::MRMSrcMem:
1208   case X86II::MRMSrcMemCC:
1209     REX |= isREXExtendedReg(MI, CurOp++) << 2;                        // REX.R
1210     REX |= isREXExtendedReg(MI, MemOperand + X86::AddrBaseReg) << 0;  // REX.B
1211     REX |= isREXExtendedReg(MI, MemOperand + X86::AddrIndexReg) << 1; // REX.X
1212     CurOp += X86::AddrNumOperands;
1213     break;
1214   case X86II::MRMDestReg:
1215     REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1216     REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1217     break;
1218   case X86II::MRMDestMem:
1219     REX |= isREXExtendedReg(MI, MemOperand + X86::AddrBaseReg) << 0;  // REX.B
1220     REX |= isREXExtendedReg(MI, MemOperand + X86::AddrIndexReg) << 1; // REX.X
1221     CurOp += X86::AddrNumOperands;
1222     REX |= isREXExtendedReg(MI, CurOp++) << 2; // REX.R
1223     break;
1224   case X86II::MRMXmCC:
1225   case X86II::MRMXm:
1226   case X86II::MRM0m:
1227   case X86II::MRM1m:
1228   case X86II::MRM2m:
1229   case X86II::MRM3m:
1230   case X86II::MRM4m:
1231   case X86II::MRM5m:
1232   case X86II::MRM6m:
1233   case X86II::MRM7m:
1234     REX |= isREXExtendedReg(MI, MemOperand + X86::AddrBaseReg) << 0;  // REX.B
1235     REX |= isREXExtendedReg(MI, MemOperand + X86::AddrIndexReg) << 1; // REX.X
1236     break;
1237   case X86II::MRMXrCC:
1238   case X86II::MRMXr:
1239   case X86II::MRM0r:
1240   case X86II::MRM1r:
1241   case X86II::MRM2r:
1242   case X86II::MRM3r:
1243   case X86II::MRM4r:
1244   case X86II::MRM5r:
1245   case X86II::MRM6r:
1246   case X86II::MRM7r:
1247     REX |= isREXExtendedReg(MI, CurOp++) << 0; // REX.B
1248     break;
1249   }
1250   if (REX && UsesHighByteReg)
1251     report_fatal_error(
1252         "Cannot encode high byte register in REX-prefixed instruction");
1253 
1254   return REX;
1255 }
1256 
1257 /// Emit segment override opcode prefix as needed.
1258 void X86MCCodeEmitter::emitSegmentOverridePrefix(unsigned &CurByte,
1259                                                  unsigned SegOperand,
1260                                                  const MCInst &MI,
1261                                                  raw_ostream &OS) const {
1262   // Check for explicit segment override on memory operand.
1263   switch (MI.getOperand(SegOperand).getReg()) {
1264   default:
1265     llvm_unreachable("Unknown segment register!");
1266   case 0:
1267     break;
1268   case X86::CS:
1269     emitByte(0x2E, CurByte, OS);
1270     break;
1271   case X86::SS:
1272     emitByte(0x36, CurByte, OS);
1273     break;
1274   case X86::DS:
1275     emitByte(0x3E, CurByte, OS);
1276     break;
1277   case X86::ES:
1278     emitByte(0x26, CurByte, OS);
1279     break;
1280   case X86::FS:
1281     emitByte(0x64, CurByte, OS);
1282     break;
1283   case X86::GS:
1284     emitByte(0x65, CurByte, OS);
1285     break;
1286   }
1287 }
1288 
1289 /// Emit all instruction prefixes prior to the opcode.
1290 ///
1291 /// \param MemOperand the operand # of the start of a memory operand if present.
1292 /// If not present, it is -1.
1293 ///
1294 /// \returns true if a REX prefix was used.
1295 bool X86MCCodeEmitter::emitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
1296                                         int MemOperand, const MCInst &MI,
1297                                         const MCInstrDesc &Desc,
1298                                         const MCSubtargetInfo &STI,
1299                                         raw_ostream &OS) const {
1300   bool Ret = false;
1301   // Emit the operand size opcode prefix as needed.
1302   if ((TSFlags & X86II::OpSizeMask) ==
1303       (STI.hasFeature(X86::Mode16Bit) ? X86II::OpSize32 : X86II::OpSize16))
1304     emitByte(0x66, CurByte, OS);
1305 
1306   // Emit the LOCK opcode prefix.
1307   if (TSFlags & X86II::LOCK || MI.getFlags() & X86::IP_HAS_LOCK)
1308     emitByte(0xF0, CurByte, OS);
1309 
1310   // Emit the NOTRACK opcode prefix.
1311   if (TSFlags & X86II::NOTRACK || MI.getFlags() & X86::IP_HAS_NOTRACK)
1312     emitByte(0x3E, CurByte, OS);
1313 
1314   switch (TSFlags & X86II::OpPrefixMask) {
1315   case X86II::PD: // 66
1316     emitByte(0x66, CurByte, OS);
1317     break;
1318   case X86II::XS: // F3
1319     emitByte(0xF3, CurByte, OS);
1320     break;
1321   case X86II::XD: // F2
1322     emitByte(0xF2, CurByte, OS);
1323     break;
1324   }
1325 
1326   // Handle REX prefix.
1327   // FIXME: Can this come before F2 etc to simplify emission?
1328   if (STI.hasFeature(X86::Mode64Bit)) {
1329     if (uint8_t REX = determineREXPrefix(MI, TSFlags, MemOperand, Desc)) {
1330       emitByte(0x40 | REX, CurByte, OS);
1331       Ret = true;
1332     }
1333   } else {
1334     assert(!(TSFlags & X86II::REX_W) && "REX.W requires 64bit mode.");
1335   }
1336 
1337   // 0x0F escape code must be emitted just before the opcode.
1338   switch (TSFlags & X86II::OpMapMask) {
1339   case X86II::TB:        // Two-byte opcode map
1340   case X86II::T8:        // 0F 38
1341   case X86II::TA:        // 0F 3A
1342   case X86II::ThreeDNow: // 0F 0F, second 0F emitted by caller.
1343     emitByte(0x0F, CurByte, OS);
1344     break;
1345   }
1346 
1347   switch (TSFlags & X86II::OpMapMask) {
1348   case X86II::T8: // 0F 38
1349     emitByte(0x38, CurByte, OS);
1350     break;
1351   case X86II::TA: // 0F 3A
1352     emitByte(0x3A, CurByte, OS);
1353     break;
1354   }
1355   return Ret;
1356 }
1357 
1358 void X86MCCodeEmitter::emitPrefix(const MCInst &MI, raw_ostream &OS,
1359                                   const MCSubtargetInfo &STI) const {
1360   unsigned Opcode = MI.getOpcode();
1361   const MCInstrDesc &Desc = MCII.get(Opcode);
1362   uint64_t TSFlags = Desc.TSFlags;
1363 
1364   // Pseudo instructions don't get encoded.
1365   if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
1366     return;
1367 
1368   unsigned CurOp = X86II::getOperandBias(Desc);
1369 
1370   // Keep track of the current byte being emitted.
1371   unsigned CurByte = 0;
1372 
1373   bool Rex = false;
1374   emitPrefixImpl(TSFlags, CurOp, CurByte, Rex, MI, Desc, STI, OS);
1375 }
1376 
1377 void X86MCCodeEmitter::encodeInstruction(const MCInst &MI, raw_ostream &OS,
1378                                          SmallVectorImpl<MCFixup> &Fixups,
1379                                          const MCSubtargetInfo &STI) const {
1380   unsigned Opcode = MI.getOpcode();
1381   const MCInstrDesc &Desc = MCII.get(Opcode);
1382   uint64_t TSFlags = Desc.TSFlags;
1383 
1384   // Pseudo instructions don't get encoded.
1385   if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
1386     return;
1387 
1388   unsigned NumOps = Desc.getNumOperands();
1389   unsigned CurOp = X86II::getOperandBias(Desc);
1390 
1391   // Keep track of the current byte being emitted.
1392   unsigned CurByte = 0;
1393 
1394   bool Rex = false;
1395   emitPrefixImpl(TSFlags, CurOp, CurByte, Rex, MI, Desc, STI, OS);
1396 
1397   // It uses the VEX.VVVV field?
1398   bool HasVEX_4V = TSFlags & X86II::VEX_4V;
1399   bool HasVEX_I8Reg = (TSFlags & X86II::ImmMask) == X86II::Imm8Reg;
1400 
1401   // It uses the EVEX.aaa field?
1402   bool HasEVEX_K = TSFlags & X86II::EVEX_K;
1403   bool HasEVEX_RC = TSFlags & X86II::EVEX_RC;
1404 
1405   // Used if a register is encoded in 7:4 of immediate.
1406   unsigned I8RegNum = 0;
1407 
1408   uint8_t BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
1409 
1410   if ((TSFlags & X86II::OpMapMask) == X86II::ThreeDNow)
1411     BaseOpcode = 0x0F; // Weird 3DNow! encoding.
1412 
1413   unsigned OpcodeOffset = 0;
1414 
1415   uint64_t Form = TSFlags & X86II::FormMask;
1416   switch (Form) {
1417   default:
1418     errs() << "FORM: " << Form << "\n";
1419     llvm_unreachable("Unknown FormMask value in X86MCCodeEmitter!");
1420   case X86II::Pseudo:
1421     llvm_unreachable("Pseudo instruction shouldn't be emitted");
1422   case X86II::RawFrmDstSrc:
1423   case X86II::RawFrmSrc:
1424   case X86II::RawFrmDst:
1425     emitByte(BaseOpcode, CurByte, OS);
1426     break;
1427   case X86II::AddCCFrm: {
1428     // This will be added to the opcode in the fallthrough.
1429     OpcodeOffset = MI.getOperand(NumOps - 1).getImm();
1430     assert(OpcodeOffset < 16 && "Unexpected opcode offset!");
1431     --NumOps; // Drop the operand from the end.
1432     LLVM_FALLTHROUGH;
1433   case X86II::RawFrm:
1434     emitByte(BaseOpcode + OpcodeOffset, CurByte, OS);
1435 
1436     if (!STI.hasFeature(X86::Mode64Bit) || !isPCRel32Branch(MI, MCII))
1437       break;
1438 
1439     const MCOperand &Op = MI.getOperand(CurOp++);
1440     emitImmediate(Op, MI.getLoc(), X86II::getSizeOfImm(TSFlags),
1441                   MCFixupKind(X86::reloc_branch_4byte_pcrel), CurByte, OS,
1442                   Fixups);
1443     break;
1444   }
1445   case X86II::RawFrmMemOffs:
1446     emitByte(BaseOpcode, CurByte, OS);
1447     emitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1448                   X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1449                   CurByte, OS, Fixups);
1450     ++CurOp; // skip segment operand
1451     break;
1452   case X86II::RawFrmImm8:
1453     emitByte(BaseOpcode, CurByte, OS);
1454     emitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1455                   X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1456                   CurByte, OS, Fixups);
1457     emitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 1, FK_Data_1, CurByte,
1458                   OS, Fixups);
1459     break;
1460   case X86II::RawFrmImm16:
1461     emitByte(BaseOpcode, CurByte, OS);
1462     emitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1463                   X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1464                   CurByte, OS, Fixups);
1465     emitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 2, FK_Data_2, CurByte,
1466                   OS, Fixups);
1467     break;
1468 
1469   case X86II::AddRegFrm:
1470     emitByte(BaseOpcode + getX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
1471     break;
1472 
1473   case X86II::MRMDestReg: {
1474     emitByte(BaseOpcode, CurByte, OS);
1475     unsigned SrcRegNum = CurOp + 1;
1476 
1477     if (HasEVEX_K) // Skip writemask
1478       ++SrcRegNum;
1479 
1480     if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1481       ++SrcRegNum;
1482 
1483     emitRegModRMByte(MI.getOperand(CurOp),
1484                      getX86RegNum(MI.getOperand(SrcRegNum)), CurByte, OS);
1485     CurOp = SrcRegNum + 1;
1486     break;
1487   }
1488   case X86II::MRMDestMem: {
1489     emitByte(BaseOpcode, CurByte, OS);
1490     unsigned SrcRegNum = CurOp + X86::AddrNumOperands;
1491 
1492     if (HasEVEX_K) // Skip writemask
1493       ++SrcRegNum;
1494 
1495     if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1496       ++SrcRegNum;
1497 
1498     emitMemModRMByte(MI, CurOp, getX86RegNum(MI.getOperand(SrcRegNum)), TSFlags,
1499                      Rex, CurByte, OS, Fixups, STI);
1500     CurOp = SrcRegNum + 1;
1501     break;
1502   }
1503   case X86II::MRMSrcReg: {
1504     emitByte(BaseOpcode, CurByte, OS);
1505     unsigned SrcRegNum = CurOp + 1;
1506 
1507     if (HasEVEX_K) // Skip writemask
1508       ++SrcRegNum;
1509 
1510     if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1511       ++SrcRegNum;
1512 
1513     emitRegModRMByte(MI.getOperand(SrcRegNum),
1514                      getX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1515     CurOp = SrcRegNum + 1;
1516     if (HasVEX_I8Reg)
1517       I8RegNum = getX86RegEncoding(MI, CurOp++);
1518     // do not count the rounding control operand
1519     if (HasEVEX_RC)
1520       --NumOps;
1521     break;
1522   }
1523   case X86II::MRMSrcReg4VOp3: {
1524     emitByte(BaseOpcode, CurByte, OS);
1525     unsigned SrcRegNum = CurOp + 1;
1526 
1527     emitRegModRMByte(MI.getOperand(SrcRegNum),
1528                      getX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1529     CurOp = SrcRegNum + 1;
1530     ++CurOp; // Encoded in VEX.VVVV
1531     break;
1532   }
1533   case X86II::MRMSrcRegOp4: {
1534     emitByte(BaseOpcode, CurByte, OS);
1535     unsigned SrcRegNum = CurOp + 1;
1536 
1537     // Skip 1st src (which is encoded in VEX_VVVV)
1538     ++SrcRegNum;
1539 
1540     // Capture 2nd src (which is encoded in Imm[7:4])
1541     assert(HasVEX_I8Reg && "MRMSrcRegOp4 should imply VEX_I8Reg");
1542     I8RegNum = getX86RegEncoding(MI, SrcRegNum++);
1543 
1544     emitRegModRMByte(MI.getOperand(SrcRegNum),
1545                      getX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1546     CurOp = SrcRegNum + 1;
1547     break;
1548   }
1549   case X86II::MRMSrcRegCC: {
1550     unsigned FirstOp = CurOp++;
1551     unsigned SecondOp = CurOp++;
1552 
1553     unsigned CC = MI.getOperand(CurOp++).getImm();
1554     emitByte(BaseOpcode + CC, CurByte, OS);
1555 
1556     emitRegModRMByte(MI.getOperand(SecondOp),
1557                      getX86RegNum(MI.getOperand(FirstOp)), CurByte, OS);
1558     break;
1559   }
1560   case X86II::MRMSrcMem: {
1561     unsigned FirstMemOp = CurOp + 1;
1562 
1563     if (HasEVEX_K) // Skip writemask
1564       ++FirstMemOp;
1565 
1566     if (HasVEX_4V)
1567       ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
1568 
1569     emitByte(BaseOpcode, CurByte, OS);
1570 
1571     emitMemModRMByte(MI, FirstMemOp, getX86RegNum(MI.getOperand(CurOp)),
1572                      TSFlags, Rex, CurByte, OS, Fixups, STI);
1573     CurOp = FirstMemOp + X86::AddrNumOperands;
1574     if (HasVEX_I8Reg)
1575       I8RegNum = getX86RegEncoding(MI, CurOp++);
1576     break;
1577   }
1578   case X86II::MRMSrcMem4VOp3: {
1579     unsigned FirstMemOp = CurOp + 1;
1580 
1581     emitByte(BaseOpcode, CurByte, OS);
1582 
1583     emitMemModRMByte(MI, FirstMemOp, getX86RegNum(MI.getOperand(CurOp)),
1584                      TSFlags, Rex, CurByte, OS, Fixups, STI);
1585     CurOp = FirstMemOp + X86::AddrNumOperands;
1586     ++CurOp; // Encoded in VEX.VVVV.
1587     break;
1588   }
1589   case X86II::MRMSrcMemOp4: {
1590     unsigned FirstMemOp = CurOp + 1;
1591 
1592     ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
1593 
1594     // Capture second register source (encoded in Imm[7:4])
1595     assert(HasVEX_I8Reg && "MRMSrcRegOp4 should imply VEX_I8Reg");
1596     I8RegNum = getX86RegEncoding(MI, FirstMemOp++);
1597 
1598     emitByte(BaseOpcode, CurByte, OS);
1599 
1600     emitMemModRMByte(MI, FirstMemOp, getX86RegNum(MI.getOperand(CurOp)),
1601                      TSFlags, Rex, CurByte, OS, Fixups, STI);
1602     CurOp = FirstMemOp + X86::AddrNumOperands;
1603     break;
1604   }
1605   case X86II::MRMSrcMemCC: {
1606     unsigned RegOp = CurOp++;
1607     unsigned FirstMemOp = CurOp;
1608     CurOp = FirstMemOp + X86::AddrNumOperands;
1609 
1610     unsigned CC = MI.getOperand(CurOp++).getImm();
1611     emitByte(BaseOpcode + CC, CurByte, OS);
1612 
1613     emitMemModRMByte(MI, FirstMemOp, getX86RegNum(MI.getOperand(RegOp)),
1614                      TSFlags, Rex, CurByte, OS, Fixups, STI);
1615     break;
1616   }
1617 
1618   case X86II::MRMXrCC: {
1619     unsigned RegOp = CurOp++;
1620 
1621     unsigned CC = MI.getOperand(CurOp++).getImm();
1622     emitByte(BaseOpcode + CC, CurByte, OS);
1623     emitRegModRMByte(MI.getOperand(RegOp), 0, CurByte, OS);
1624     break;
1625   }
1626 
1627   case X86II::MRMXr:
1628   case X86II::MRM0r:
1629   case X86II::MRM1r:
1630   case X86II::MRM2r:
1631   case X86II::MRM3r:
1632   case X86II::MRM4r:
1633   case X86II::MRM5r:
1634   case X86II::MRM6r:
1635   case X86II::MRM7r:
1636     if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1637       ++CurOp;
1638     if (HasEVEX_K) // Skip writemask
1639       ++CurOp;
1640     emitByte(BaseOpcode, CurByte, OS);
1641     emitRegModRMByte(MI.getOperand(CurOp++),
1642                      (Form == X86II::MRMXr) ? 0 : Form - X86II::MRM0r, CurByte,
1643                      OS);
1644     break;
1645 
1646   case X86II::MRMXmCC: {
1647     unsigned FirstMemOp = CurOp;
1648     CurOp = FirstMemOp + X86::AddrNumOperands;
1649 
1650     unsigned CC = MI.getOperand(CurOp++).getImm();
1651     emitByte(BaseOpcode + CC, CurByte, OS);
1652 
1653     emitMemModRMByte(MI, FirstMemOp, 0, TSFlags, Rex, CurByte, OS, Fixups, STI);
1654     break;
1655   }
1656 
1657   case X86II::MRMXm:
1658   case X86II::MRM0m:
1659   case X86II::MRM1m:
1660   case X86II::MRM2m:
1661   case X86II::MRM3m:
1662   case X86II::MRM4m:
1663   case X86II::MRM5m:
1664   case X86II::MRM6m:
1665   case X86II::MRM7m:
1666     if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1667       ++CurOp;
1668     if (HasEVEX_K) // Skip writemask
1669       ++CurOp;
1670     emitByte(BaseOpcode, CurByte, OS);
1671     emitMemModRMByte(MI, CurOp,
1672                      (Form == X86II::MRMXm) ? 0 : Form - X86II::MRM0m, TSFlags,
1673                      Rex, CurByte, OS, Fixups, STI);
1674     CurOp += X86::AddrNumOperands;
1675     break;
1676 
1677   case X86II::MRM_C0:
1678   case X86II::MRM_C1:
1679   case X86II::MRM_C2:
1680   case X86II::MRM_C3:
1681   case X86II::MRM_C4:
1682   case X86II::MRM_C5:
1683   case X86II::MRM_C6:
1684   case X86II::MRM_C7:
1685   case X86II::MRM_C8:
1686   case X86II::MRM_C9:
1687   case X86II::MRM_CA:
1688   case X86II::MRM_CB:
1689   case X86II::MRM_CC:
1690   case X86II::MRM_CD:
1691   case X86II::MRM_CE:
1692   case X86II::MRM_CF:
1693   case X86II::MRM_D0:
1694   case X86II::MRM_D1:
1695   case X86II::MRM_D2:
1696   case X86II::MRM_D3:
1697   case X86II::MRM_D4:
1698   case X86II::MRM_D5:
1699   case X86II::MRM_D6:
1700   case X86II::MRM_D7:
1701   case X86II::MRM_D8:
1702   case X86II::MRM_D9:
1703   case X86II::MRM_DA:
1704   case X86II::MRM_DB:
1705   case X86II::MRM_DC:
1706   case X86II::MRM_DD:
1707   case X86II::MRM_DE:
1708   case X86II::MRM_DF:
1709   case X86II::MRM_E0:
1710   case X86II::MRM_E1:
1711   case X86II::MRM_E2:
1712   case X86II::MRM_E3:
1713   case X86II::MRM_E4:
1714   case X86II::MRM_E5:
1715   case X86II::MRM_E6:
1716   case X86II::MRM_E7:
1717   case X86II::MRM_E8:
1718   case X86II::MRM_E9:
1719   case X86II::MRM_EA:
1720   case X86II::MRM_EB:
1721   case X86II::MRM_EC:
1722   case X86II::MRM_ED:
1723   case X86II::MRM_EE:
1724   case X86II::MRM_EF:
1725   case X86II::MRM_F0:
1726   case X86II::MRM_F1:
1727   case X86II::MRM_F2:
1728   case X86II::MRM_F3:
1729   case X86II::MRM_F4:
1730   case X86II::MRM_F5:
1731   case X86II::MRM_F6:
1732   case X86II::MRM_F7:
1733   case X86II::MRM_F8:
1734   case X86II::MRM_F9:
1735   case X86II::MRM_FA:
1736   case X86II::MRM_FB:
1737   case X86II::MRM_FC:
1738   case X86II::MRM_FD:
1739   case X86II::MRM_FE:
1740   case X86II::MRM_FF:
1741     emitByte(BaseOpcode, CurByte, OS);
1742     emitByte(0xC0 + Form - X86II::MRM_C0, CurByte, OS);
1743     break;
1744   }
1745 
1746   if (HasVEX_I8Reg) {
1747     // The last source register of a 4 operand instruction in AVX is encoded
1748     // in bits[7:4] of a immediate byte.
1749     assert(I8RegNum < 16 && "Register encoding out of range");
1750     I8RegNum <<= 4;
1751     if (CurOp != NumOps) {
1752       unsigned Val = MI.getOperand(CurOp++).getImm();
1753       assert(Val < 16 && "Immediate operand value out of range");
1754       I8RegNum |= Val;
1755     }
1756     emitImmediate(MCOperand::createImm(I8RegNum), MI.getLoc(), 1, FK_Data_1,
1757                   CurByte, OS, Fixups);
1758   } else {
1759     // If there is a remaining operand, it must be a trailing immediate. Emit it
1760     // according to the right size for the instruction. Some instructions
1761     // (SSE4a extrq and insertq) have two trailing immediates.
1762     while (CurOp != NumOps && NumOps - CurOp <= 2) {
1763       emitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1764                     X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1765                     CurByte, OS, Fixups);
1766     }
1767   }
1768 
1769   if ((TSFlags & X86II::OpMapMask) == X86II::ThreeDNow)
1770     emitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS);
1771 
1772 #ifndef NDEBUG
1773   // FIXME: Verify.
1774   if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
1775     errs() << "Cannot encode all operands of: ";
1776     MI.dump();
1777     errs() << '\n';
1778     abort();
1779   }
1780 #endif
1781 }
1782 
1783 MCCodeEmitter *llvm::createX86MCCodeEmitter(const MCInstrInfo &MCII,
1784                                             const MCRegisterInfo &MRI,
1785                                             MCContext &Ctx) {
1786   return new X86MCCodeEmitter(MCII, Ctx);
1787 }
1788