1 //===-- X86MCCodeEmitter.cpp - Convert X86 code to machine code -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the X86MCCodeEmitter class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "mccodeemitter"
15 #include "MCTargetDesc/X86MCTargetDesc.h"
16 #include "MCTargetDesc/X86BaseInfo.h"
17 #include "MCTargetDesc/X86FixupKinds.h"
18 #include "llvm/MC/MCCodeEmitter.h"
19 #include "llvm/MC/MCContext.h"
20 #include "llvm/MC/MCExpr.h"
21 #include "llvm/MC/MCInst.h"
22 #include "llvm/MC/MCInstrInfo.h"
23 #include "llvm/MC/MCRegisterInfo.h"
24 #include "llvm/MC/MCSubtargetInfo.h"
25 #include "llvm/MC/MCSymbol.h"
26 #include "llvm/Support/raw_ostream.h"
31 class X86MCCodeEmitter : public MCCodeEmitter {
32 X86MCCodeEmitter(const X86MCCodeEmitter &) LLVM_DELETED_FUNCTION;
33 void operator=(const X86MCCodeEmitter &) LLVM_DELETED_FUNCTION;
34 const MCInstrInfo &MCII;
35 const MCSubtargetInfo &STI;
38 X86MCCodeEmitter(const MCInstrInfo &mcii, const MCSubtargetInfo &sti,
40 : MCII(mcii), STI(sti), Ctx(ctx) {
43 ~X86MCCodeEmitter() {}
45 bool is64BitMode() const {
46 // FIXME: Can tablegen auto-generate this?
47 return (STI.getFeatureBits() & X86::Mode64Bit) != 0;
50 bool is32BitMode() const {
51 // FIXME: Can tablegen auto-generate this?
52 return (STI.getFeatureBits() & X86::Mode64Bit) == 0;
55 unsigned GetX86RegNum(const MCOperand &MO) const {
56 return Ctx.getRegisterInfo()->getEncodingValue(MO.getReg()) & 0x7;
59 // On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the range
60 // 0-7 and the difference between the 2 groups is given by the REX prefix.
61 // In the VEX prefix, registers are seen sequencially from 0-15 and encoded
62 // in 1's complement form, example:
64 // ModRM field => XMM9 => 1
65 // VEX.VVVV => XMM9 => ~9
67 // See table 4-35 of Intel AVX Programming Reference for details.
68 unsigned char getVEXRegisterEncoding(const MCInst &MI,
69 unsigned OpNum) const {
70 unsigned SrcReg = MI.getOperand(OpNum).getReg();
71 unsigned SrcRegNum = GetX86RegNum(MI.getOperand(OpNum));
72 if (X86II::isX86_64ExtendedReg(SrcReg))
75 // The registers represented through VEX_VVVV should
76 // be encoded in 1's complement form.
77 return (~SrcRegNum) & 0xf;
80 unsigned char getWriteMaskRegisterEncoding(const MCInst &MI,
81 unsigned OpNum) const {
82 assert(X86::K0 != MI.getOperand(OpNum).getReg() &&
83 "Invalid mask register as write-mask!");
84 unsigned MaskRegNum = GetX86RegNum(MI.getOperand(OpNum));
88 void EmitByte(unsigned char C, unsigned &CurByte, raw_ostream &OS) const {
93 void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
94 raw_ostream &OS) const {
95 // Output the constant in little endian byte order.
96 for (unsigned i = 0; i != Size; ++i) {
97 EmitByte(Val & 255, CurByte, OS);
102 void EmitImmediate(const MCOperand &Disp, SMLoc Loc,
103 unsigned ImmSize, MCFixupKind FixupKind,
104 unsigned &CurByte, raw_ostream &OS,
105 SmallVectorImpl<MCFixup> &Fixups,
106 int ImmOffset = 0) const;
108 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
110 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
111 return RM | (RegOpcode << 3) | (Mod << 6);
114 void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
115 unsigned &CurByte, raw_ostream &OS) const {
116 EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS);
119 void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base,
120 unsigned &CurByte, raw_ostream &OS) const {
121 // SIB byte is in the same format as the ModRMByte.
122 EmitByte(ModRMByte(SS, Index, Base), CurByte, OS);
126 void EmitMemModRMByte(const MCInst &MI, unsigned Op,
127 unsigned RegOpcodeField,
128 uint64_t TSFlags, unsigned &CurByte, raw_ostream &OS,
129 SmallVectorImpl<MCFixup> &Fixups) const;
131 void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
132 SmallVectorImpl<MCFixup> &Fixups) const;
134 void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
135 const MCInst &MI, const MCInstrDesc &Desc,
136 raw_ostream &OS) const;
138 void EmitSegmentOverridePrefix(uint64_t TSFlags, unsigned &CurByte,
139 int MemOperand, const MCInst &MI,
140 raw_ostream &OS) const;
142 void EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
143 const MCInst &MI, const MCInstrDesc &Desc,
144 raw_ostream &OS) const;
147 } // end anonymous namespace
150 MCCodeEmitter *llvm::createX86MCCodeEmitter(const MCInstrInfo &MCII,
151 const MCRegisterInfo &MRI,
152 const MCSubtargetInfo &STI,
154 return new X86MCCodeEmitter(MCII, STI, Ctx);
157 /// isDisp8 - Return true if this signed displacement fits in a 8-bit
158 /// sign-extended field.
159 static bool isDisp8(int Value) {
160 return Value == (signed char)Value;
163 /// isCDisp8 - Return true if this signed displacement fits in a 8-bit
164 /// compressed dispacement field.
165 static bool isCDisp8(uint64_t TSFlags, int Value, int& CValue) {
166 assert(((TSFlags >> X86II::VEXShift) & X86II::EVEX) &&
167 "Compressed 8-bit displacement is only valid for EVEX inst.");
169 unsigned CD8E = (TSFlags >> X86II::EVEX_CD8EShift) & X86II::EVEX_CD8EMask;
170 unsigned CD8V = (TSFlags >> X86II::EVEX_CD8VShift) & X86II::EVEX_CD8VMask;
172 if (CD8V == 0 && CD8E == 0) {
174 return isDisp8(Value);
177 unsigned MemObjSize = 1U << CD8E;
179 // Fixed vector length
180 MemObjSize *= 1U << (CD8V & 0x3);
182 // Modified vector length
183 bool EVEX_b = (TSFlags >> X86II::VEXShift) & X86II::EVEX_B;
185 unsigned EVEX_LL = ((TSFlags >> X86II::VEXShift) & X86II::VEX_L) ? 1 : 0;
186 EVEX_LL += ((TSFlags >> X86II::VEXShift) & X86II::EVEX_L2) ? 2 : 0;
187 assert(EVEX_LL < 3 && "");
189 unsigned NumElems = (1U << (EVEX_LL + 4)) / MemObjSize;
190 NumElems /= 1U << (CD8V & 0x3);
192 MemObjSize *= NumElems;
196 unsigned MemObjMask = MemObjSize - 1;
197 assert((MemObjSize & MemObjMask) == 0 && "Invalid memory object size.");
199 if (Value & MemObjMask) // Unaligned offset
202 bool Ret = (Value == (signed char)Value);
209 /// getImmFixupKind - Return the appropriate fixup kind to use for an immediate
210 /// in an instruction with the specified TSFlags.
211 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
212 unsigned Size = X86II::getSizeOfImm(TSFlags);
213 bool isPCRel = X86II::isImmPCRel(TSFlags);
215 return MCFixup::getKindForSize(Size, isPCRel);
218 /// Is32BitMemOperand - Return true if the specified instruction has
219 /// a 32-bit memory operand. Op specifies the operand # of the memoperand.
220 static bool Is32BitMemOperand(const MCInst &MI, unsigned Op) {
221 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
222 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
224 if ((BaseReg.getReg() != 0 &&
225 X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) ||
226 (IndexReg.getReg() != 0 &&
227 X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg())))
232 /// Is64BitMemOperand - Return true if the specified instruction has
233 /// a 64-bit memory operand. Op specifies the operand # of the memoperand.
235 static bool Is64BitMemOperand(const MCInst &MI, unsigned Op) {
236 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
237 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
239 if ((BaseReg.getReg() != 0 &&
240 X86MCRegisterClasses[X86::GR64RegClassID].contains(BaseReg.getReg())) ||
241 (IndexReg.getReg() != 0 &&
242 X86MCRegisterClasses[X86::GR64RegClassID].contains(IndexReg.getReg())))
248 /// Is16BitMemOperand - Return true if the specified instruction has
249 /// a 16-bit memory operand. Op specifies the operand # of the memoperand.
250 static bool Is16BitMemOperand(const MCInst &MI, unsigned Op) {
251 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
252 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
254 if ((BaseReg.getReg() != 0 &&
255 X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) ||
256 (IndexReg.getReg() != 0 &&
257 X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg())))
262 /// StartsWithGlobalOffsetTable - Check if this expression starts with
263 /// _GLOBAL_OFFSET_TABLE_ and if it is of the form
264 /// _GLOBAL_OFFSET_TABLE_-symbol. This is needed to support PIC on ELF
265 /// i386 as _GLOBAL_OFFSET_TABLE_ is magical. We check only simple case that
266 /// are know to be used: _GLOBAL_OFFSET_TABLE_ by itself or at the start
267 /// of a binary expression.
268 enum GlobalOffsetTableExprKind {
273 static GlobalOffsetTableExprKind
274 StartsWithGlobalOffsetTable(const MCExpr *Expr) {
275 const MCExpr *RHS = 0;
276 if (Expr->getKind() == MCExpr::Binary) {
277 const MCBinaryExpr *BE = static_cast<const MCBinaryExpr *>(Expr);
282 if (Expr->getKind() != MCExpr::SymbolRef)
285 const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
286 const MCSymbol &S = Ref->getSymbol();
287 if (S.getName() != "_GLOBAL_OFFSET_TABLE_")
289 if (RHS && RHS->getKind() == MCExpr::SymbolRef)
294 static bool HasSecRelSymbolRef(const MCExpr *Expr) {
295 if (Expr->getKind() == MCExpr::SymbolRef) {
296 const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
297 return Ref->getKind() == MCSymbolRefExpr::VK_SECREL;
302 void X86MCCodeEmitter::
303 EmitImmediate(const MCOperand &DispOp, SMLoc Loc, unsigned Size,
304 MCFixupKind FixupKind, unsigned &CurByte, raw_ostream &OS,
305 SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
306 const MCExpr *Expr = NULL;
307 if (DispOp.isImm()) {
308 // If this is a simple integer displacement that doesn't require a
309 // relocation, emit it now.
310 if (FixupKind != FK_PCRel_1 &&
311 FixupKind != FK_PCRel_2 &&
312 FixupKind != FK_PCRel_4) {
313 EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
316 Expr = MCConstantExpr::Create(DispOp.getImm(), Ctx);
318 Expr = DispOp.getExpr();
321 // If we have an immoffset, add it to the expression.
322 if ((FixupKind == FK_Data_4 ||
323 FixupKind == FK_Data_8 ||
324 FixupKind == MCFixupKind(X86::reloc_signed_4byte))) {
325 GlobalOffsetTableExprKind Kind = StartsWithGlobalOffsetTable(Expr);
326 if (Kind != GOT_None) {
327 assert(ImmOffset == 0);
329 FixupKind = MCFixupKind(X86::reloc_global_offset_table);
330 if (Kind == GOT_Normal)
332 } else if (Expr->getKind() == MCExpr::SymbolRef) {
333 if (HasSecRelSymbolRef(Expr)) {
334 FixupKind = MCFixupKind(FK_SecRel_4);
336 } else if (Expr->getKind() == MCExpr::Binary) {
337 const MCBinaryExpr *Bin = static_cast<const MCBinaryExpr*>(Expr);
338 if (HasSecRelSymbolRef(Bin->getLHS())
339 || HasSecRelSymbolRef(Bin->getRHS())) {
340 FixupKind = MCFixupKind(FK_SecRel_4);
345 // If the fixup is pc-relative, we need to bias the value to be relative to
346 // the start of the field, not the end of the field.
347 if (FixupKind == FK_PCRel_4 ||
348 FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
349 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load))
351 if (FixupKind == FK_PCRel_2)
353 if (FixupKind == FK_PCRel_1)
357 Expr = MCBinaryExpr::CreateAdd(Expr, MCConstantExpr::Create(ImmOffset, Ctx),
360 // Emit a symbolic constant as a fixup and 4 zeros.
361 Fixups.push_back(MCFixup::Create(CurByte, Expr, FixupKind, Loc));
362 EmitConstant(0, Size, CurByte, OS);
365 void X86MCCodeEmitter::EmitMemModRMByte(const MCInst &MI, unsigned Op,
366 unsigned RegOpcodeField,
367 uint64_t TSFlags, unsigned &CurByte,
369 SmallVectorImpl<MCFixup> &Fixups) const{
370 const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
371 const MCOperand &Base = MI.getOperand(Op+X86::AddrBaseReg);
372 const MCOperand &Scale = MI.getOperand(Op+X86::AddrScaleAmt);
373 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
374 unsigned BaseReg = Base.getReg();
375 bool HasEVEX = (TSFlags >> X86II::VEXShift) & X86II::EVEX;
377 // Handle %rip relative addressing.
378 if (BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
379 assert(is64BitMode() && "Rip-relative addressing requires 64-bit mode");
380 assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
381 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
383 unsigned FixupKind = X86::reloc_riprel_4byte;
385 // movq loads are handled with a special relocation form which allows the
386 // linker to eliminate some loads for GOT references which end up in the
387 // same linkage unit.
388 if (MI.getOpcode() == X86::MOV64rm)
389 FixupKind = X86::reloc_riprel_4byte_movq_load;
391 // rip-relative addressing is actually relative to the *next* instruction.
392 // Since an immediate can follow the mod/rm byte for an instruction, this
393 // means that we need to bias the immediate field of the instruction with
394 // the size of the immediate field. If we have this case, add it into the
395 // expression to emit.
396 int ImmSize = X86II::hasImm(TSFlags) ? X86II::getSizeOfImm(TSFlags) : 0;
398 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(FixupKind),
399 CurByte, OS, Fixups, -ImmSize);
403 unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
405 // 16-bit addressing forms of the ModR/M byte have a different encoding for
406 // the R/M field and are far more limited in which registers can be used.
407 if (Is16BitMemOperand(MI, Op)) {
409 // For 32-bit addressing, the row and column values in Table 2-2 are
410 // basically the same. It's AX/CX/DX/BX/SP/BP/SI/DI in that order, with
411 // some special cases. And GetX86RegNum reflects that numbering.
412 // For 16-bit addressing it's more fun, as shown in the SDM Vol 2A,
413 // Table 2-1 "16-Bit Addressing Forms with the ModR/M byte". We can only
414 // use SI/DI/BP/BX, which have "row" values 4-7 in no particular order,
415 // while values 0-3 indicate the allowed combinations (base+index) of
416 // those: 0 for BX+SI, 1 for BX+DI, 2 for BP+SI, 3 for BP+DI.
418 // R16Table[] is a lookup from the normal RegNo, to the row values from
419 // Table 2-1 for 16-bit addressing modes. Where zero means disallowed.
420 static const unsigned R16Table[] = { 0, 0, 0, 7, 0, 6, 4, 5 };
421 unsigned RMfield = R16Table[BaseRegNo];
423 assert(RMfield && "invalid 16-bit base register");
425 if (IndexReg.getReg()) {
426 unsigned IndexReg16 = R16Table[GetX86RegNum(IndexReg)];
428 assert(IndexReg16 && "invalid 16-bit index register");
429 // We must have one of SI/DI (4,5), and one of BP/BX (6,7).
430 assert(((IndexReg16 ^ RMfield) & 2) &&
431 "invalid 16-bit base/index register combination");
432 assert(Scale.getImm() == 1 &&
433 "invalid scale for 16-bit memory reference");
435 // Allow base/index to appear in either order (although GAS doesn't).
437 RMfield = (RMfield & 1) | ((7 - IndexReg16) << 1);
439 RMfield = (IndexReg16 & 1) | ((7 - RMfield) << 1);
442 if (Disp.isImm() && isDisp8(Disp.getImm())) {
443 if (Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
444 // There is no displacement; just the register.
445 EmitByte(ModRMByte(0, RegOpcodeField, RMfield), CurByte, OS);
448 // Use the [REG]+disp8 form, including for [BP] which cannot be encoded.
449 EmitByte(ModRMByte(1, RegOpcodeField, RMfield), CurByte, OS);
450 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
453 // This is the [REG]+disp16 case.
454 EmitByte(ModRMByte(2, RegOpcodeField, RMfield), CurByte, OS);
456 // There is no BaseReg; this is the plain [disp16] case.
457 EmitByte(ModRMByte(0, RegOpcodeField, 6), CurByte, OS);
460 // Emit 16-bit displacement for plain disp16 or [REG]+disp16 cases.
461 EmitImmediate(Disp, MI.getLoc(), 2, FK_Data_2, CurByte, OS, Fixups);
465 // Determine whether a SIB byte is needed.
466 // If no BaseReg, issue a RIP relative instruction only if the MCE can
467 // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
468 // 2-7) and absolute references.
470 if (// The SIB byte must be used if there is an index register.
471 IndexReg.getReg() == 0 &&
472 // The SIB byte must be used if the base is ESP/RSP/R12, all of which
473 // encode to an R/M value of 4, which indicates that a SIB byte is
475 BaseRegNo != N86::ESP &&
476 // If there is no base register and we're in 64-bit mode, we need a SIB
477 // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
478 (!is64BitMode() || BaseReg != 0)) {
480 if (BaseReg == 0) { // [disp32] in X86-32 mode
481 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
482 EmitImmediate(Disp, MI.getLoc(), 4, FK_Data_4, CurByte, OS, Fixups);
486 // If the base is not EBP/ESP and there is no displacement, use simple
487 // indirect register encoding, this handles addresses like [EAX]. The
488 // encoding for [EBP] with no displacement means [disp32] so we handle it
489 // by emitting a displacement of 0 below.
490 if (Disp.isImm() && Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
491 EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
495 // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
497 if (!HasEVEX && isDisp8(Disp.getImm())) {
498 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
499 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups);
502 // Try EVEX compressed 8-bit displacement first; if failed, fall back to
503 // 32-bit displacement.
505 if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
506 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
507 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups,
508 CDisp8 - Disp.getImm());
513 // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
514 EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
515 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(X86::reloc_signed_4byte), CurByte, OS,
520 // We need a SIB byte, so start by outputting the ModR/M byte first
521 assert(IndexReg.getReg() != X86::ESP &&
522 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
524 bool ForceDisp32 = false;
525 bool ForceDisp8 = false;
529 // If there is no base register, we emit the special case SIB byte with
530 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
531 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
533 } else if (!Disp.isImm()) {
534 // Emit the normal disp32 encoding.
535 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
537 } else if (Disp.getImm() == 0 &&
538 // Base reg can't be anything that ends up with '5' as the base
539 // reg, it is the magic [*] nomenclature that indicates no base.
540 BaseRegNo != N86::EBP) {
541 // Emit no displacement ModR/M byte
542 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
543 } else if (!HasEVEX && isDisp8(Disp.getImm())) {
544 // Emit the disp8 encoding.
545 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
546 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
547 } else if (HasEVEX && isCDisp8(TSFlags, Disp.getImm(), CDisp8)) {
548 // Emit the disp8 encoding.
549 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
550 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
551 ImmOffset = CDisp8 - Disp.getImm();
553 // Emit the normal disp32 encoding.
554 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
557 // Calculate what the SS field value should be...
558 static const unsigned SSTable[] = { ~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3 };
559 unsigned SS = SSTable[Scale.getImm()];
562 // Handle the SIB byte for the case where there is no base, see Intel
563 // Manual 2A, table 2-7. The displacement has already been output.
565 if (IndexReg.getReg())
566 IndexRegNo = GetX86RegNum(IndexReg);
567 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
569 EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
572 if (IndexReg.getReg())
573 IndexRegNo = GetX86RegNum(IndexReg);
575 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
576 EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
579 // Do we need to output a displacement?
581 EmitImmediate(Disp, MI.getLoc(), 1, FK_Data_1, CurByte, OS, Fixups, ImmOffset);
582 else if (ForceDisp32 || Disp.getImm() != 0)
583 EmitImmediate(Disp, MI.getLoc(), 4, MCFixupKind(X86::reloc_signed_4byte),
584 CurByte, OS, Fixups);
587 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
589 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
590 int MemOperand, const MCInst &MI,
591 const MCInstrDesc &Desc,
592 raw_ostream &OS) const {
593 bool HasEVEX = (TSFlags >> X86II::VEXShift) & X86II::EVEX;
594 bool HasEVEX_K = HasEVEX && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_K);
595 bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
596 bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
597 bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4;
598 bool HasEVEX_RC = false;
600 // VEX_R: opcode externsion equivalent to REX.R in
601 // 1's complement (inverted) form
603 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
604 // 0: Same as REX_R=1 (64 bit mode only)
606 unsigned char VEX_R = 0x1;
607 unsigned char EVEX_R2 = 0x1;
609 // VEX_X: equivalent to REX.X, only used when a
610 // register is used for index in SIB Byte.
612 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
613 // 0: Same as REX.X=1 (64-bit mode only)
614 unsigned char VEX_X = 0x1;
618 // 1: Same as REX_B=0 (ignored in 32-bit mode)
619 // 0: Same as REX_B=1 (64 bit mode only)
621 unsigned char VEX_B = 0x1;
623 // VEX_W: opcode specific (use like REX.W, or used for
624 // opcode extension, or ignored, depending on the opcode byte)
625 unsigned char VEX_W = 0;
627 // XOP: Use XOP prefix byte 0x8f instead of VEX.
630 // VEX_5M (VEX m-mmmmm field):
632 // 0b00000: Reserved for future use
633 // 0b00001: implied 0F leading opcode
634 // 0b00010: implied 0F 38 leading opcode bytes
635 // 0b00011: implied 0F 3A leading opcode bytes
636 // 0b00100-0b11111: Reserved for future use
637 // 0b01000: XOP map select - 08h instructions with imm byte
638 // 0b01001: XOP map select - 09h instructions with no imm byte
639 // 0b01010: XOP map select - 0Ah instructions with imm dword
640 unsigned char VEX_5M = 0x1;
642 // VEX_4V (VEX vvvv field): a register specifier
643 // (in 1's complement form) or 1111 if unused.
644 unsigned char VEX_4V = 0xf;
645 unsigned char EVEX_V2 = 0x1;
647 // VEX_L (Vector Length):
649 // 0: scalar or 128-bit vector
652 unsigned char VEX_L = 0;
653 unsigned char EVEX_L2 = 0;
655 // VEX_PP: opcode extension providing equivalent
656 // functionality of a SIMD prefix
663 unsigned char VEX_PP = 0;
666 unsigned char EVEX_U = 1; // Always '1' so far
669 unsigned char EVEX_z = 0;
672 unsigned char EVEX_b = 0;
675 unsigned char EVEX_rc = 0;
678 unsigned char EVEX_aaa = 0;
680 // Encode the operand size opcode prefix as needed.
681 if (TSFlags & X86II::OpSize)
684 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_W)
687 if ((TSFlags >> X86II::VEXShift) & X86II::XOP)
690 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_L)
692 if (HasEVEX && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_L2))
695 if (HasEVEX_K && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_Z))
698 if (HasEVEX && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_B))
701 switch (TSFlags & X86II::Op0Mask) {
702 default: llvm_unreachable("Invalid prefix!");
703 case X86II::T8: // 0F 38
706 case X86II::TA: // 0F 3A
709 case X86II::T8XS: // F3 0F 38
713 case X86II::T8XD: // F2 0F 38
717 case X86II::TAXD: // F2 0F 3A
721 case X86II::XS: // F3 0F
724 case X86II::XD: // F2 0F
736 case X86II::TB: // VEX_5M/VEX_PP already correct
741 // Classify VEX_B, VEX_4V, VEX_R, VEX_X
742 unsigned NumOps = Desc.getNumOperands();
743 unsigned RcOperand = NumOps-1;
745 if (NumOps > 1 && Desc.getOperandConstraint(1, MCOI::TIED_TO) == 0)
747 else if (NumOps > 3 && Desc.getOperandConstraint(2, MCOI::TIED_TO) == 0 &&
748 Desc.getOperandConstraint(3, MCOI::TIED_TO) == 1)
749 // Special case for AVX-512 GATHER with 2 TIED_TO operands
750 // Skip the first 2 operands: dst, mask_wb
752 else if (NumOps > 3 && Desc.getOperandConstraint(2, MCOI::TIED_TO) == 0 &&
753 Desc.getOperandConstraint(NumOps - 1, MCOI::TIED_TO) == 1)
754 // Special case for GATHER with 2 TIED_TO operands
755 // Skip the first 2 operands: dst, mask_wb
757 else if (NumOps > 2 && Desc.getOperandConstraint(NumOps - 2, MCOI::TIED_TO) == 0)
761 switch (TSFlags & X86II::FormMask) {
762 case X86II::MRMDestMem: {
763 // MRMDestMem instructions forms:
764 // MemAddr, src1(ModR/M)
765 // MemAddr, src1(VEX_4V), src2(ModR/M)
766 // MemAddr, src1(ModR/M), imm8
768 if (X86II::isX86_64ExtendedReg(MI.getOperand(MemOperand +
769 X86::AddrBaseReg).getReg()))
771 if (X86II::isX86_64ExtendedReg(MI.getOperand(MemOperand +
772 X86::AddrIndexReg).getReg()))
774 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(MemOperand +
775 X86::AddrIndexReg).getReg()))
778 CurOp += X86::AddrNumOperands;
781 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
784 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
785 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
790 const MCOperand &MO = MI.getOperand(CurOp);
792 if (X86II::isX86_64ExtendedReg(MO.getReg()))
794 if (HasEVEX && X86II::is32ExtendedReg(MO.getReg()))
799 case X86II::MRMSrcMem:
800 // MRMSrcMem instructions forms:
801 // src1(ModR/M), MemAddr
802 // src1(ModR/M), src2(VEX_4V), MemAddr
803 // src1(ModR/M), MemAddr, imm8
804 // src1(ModR/M), MemAddr, src2(VEX_I8IMM)
807 // dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
808 // dst(ModR/M.reg), src1(VEX_4V), src2(VEX_I8IMM), src3(ModR/M),
809 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
811 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
816 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
819 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
820 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
825 if (X86II::isX86_64ExtendedReg(
826 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
828 if (X86II::isX86_64ExtendedReg(
829 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
831 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(MemOperand +
832 X86::AddrIndexReg).getReg()))
836 // Instruction format for 4VOp3:
837 // src1(ModR/M), MemAddr, src3(VEX_4V)
838 // CurOp points to start of the MemoryOperand,
839 // it skips TIED_TO operands if exist, then increments past src1.
840 // CurOp + X86::AddrNumOperands will point to src3.
841 VEX_4V = getVEXRegisterEncoding(MI, CurOp+X86::AddrNumOperands);
843 case X86II::MRM0m: case X86II::MRM1m:
844 case X86II::MRM2m: case X86II::MRM3m:
845 case X86II::MRM4m: case X86II::MRM5m:
846 case X86II::MRM6m: case X86II::MRM7m: {
847 // MRM[0-9]m instructions forms:
849 // src1(VEX_4V), MemAddr
851 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
852 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
858 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
860 if (X86II::isX86_64ExtendedReg(
861 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
863 if (X86II::isX86_64ExtendedReg(
864 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
868 case X86II::MRMSrcReg:
869 // MRMSrcReg instructions forms:
870 // dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
871 // dst(ModR/M), src1(ModR/M)
872 // dst(ModR/M), src1(ModR/M), imm8
875 // dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
876 // dst(ModR/M.reg), src1(VEX_4V), src2(VEX_I8IMM), src3(ModR/M),
877 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
879 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
884 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
887 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
888 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
893 if (HasMemOp4) // Skip second register source (encoded in I8IMM)
896 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
898 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
902 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
904 assert(RcOperand >= CurOp);
905 EVEX_rc = MI.getOperand(RcOperand).getImm() & 0x3;
909 case X86II::MRMDestReg:
910 // MRMDestReg instructions forms:
911 // dst(ModR/M), src(ModR/M)
912 // dst(ModR/M), src(ModR/M), imm8
913 // dst(ModR/M), src1(VEX_4V), src2(ModR/M)
914 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
916 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
921 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
924 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
925 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
930 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
932 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
935 case X86II::MRM0r: case X86II::MRM1r:
936 case X86II::MRM2r: case X86II::MRM3r:
937 case X86II::MRM4r: case X86II::MRM5r:
938 case X86II::MRM6r: case X86II::MRM7r:
939 // MRM0r-MRM7r instructions forms:
940 // dst(VEX_4V), src(ModR/M), imm8
942 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
943 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
948 EVEX_aaa = getWriteMaskRegisterEncoding(MI, CurOp++);
950 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
952 if (HasEVEX && X86II::is32ExtendedReg(MI.getOperand(CurOp).getReg()))
959 // Emit segment override opcode prefix as needed.
960 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
963 // VEX opcode prefix can have 2 or 3 bytes
966 // +-----+ +--------------+ +-------------------+
967 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
968 // +-----+ +--------------+ +-------------------+
970 // +-----+ +-------------------+
971 // | C5h | | R | vvvv | L | pp |
972 // +-----+ +-------------------+
974 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
976 if (VEX_B && VEX_X && !VEX_W && !XOP && (VEX_5M == 1)) { // 2 byte VEX prefix
977 EmitByte(0xC5, CurByte, OS);
978 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
983 EmitByte(XOP ? 0x8F : 0xC4, CurByte, OS);
984 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
985 EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
987 // EVEX opcode prefix can have 4 bytes
989 // +-----+ +--------------+ +-------------------+ +------------------------+
990 // | 62h | | RXBR' | 00mm | | W | vvvv | U | pp | | z | L'L | b | v' | aaa |
991 // +-----+ +--------------+ +-------------------+ +------------------------+
992 assert((VEX_5M & 0x3) == VEX_5M
993 && "More than 2 significant bits in VEX.m-mmmm fields for EVEX!");
997 EmitByte(0x62, CurByte, OS);
998 EmitByte((VEX_R << 7) |
1002 VEX_5M, CurByte, OS);
1003 EmitByte((VEX_W << 7) |
1006 VEX_PP, CurByte, OS);
1008 EmitByte((EVEX_z << 7) |
1012 EVEX_aaa, CurByte, OS);
1014 EmitByte((EVEX_z << 7) |
1019 EVEX_aaa, CurByte, OS);
1023 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
1024 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
1025 /// size, and 3) use of X86-64 extended registers.
1026 static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
1027 const MCInstrDesc &Desc) {
1029 if (TSFlags & X86II::REX_W)
1030 REX |= 1 << 3; // set REX.W
1032 if (MI.getNumOperands() == 0) return REX;
1034 unsigned NumOps = MI.getNumOperands();
1035 // FIXME: MCInst should explicitize the two-addrness.
1036 bool isTwoAddr = NumOps > 1 &&
1037 Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1;
1039 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
1040 unsigned i = isTwoAddr ? 1 : 0;
1041 for (; i != NumOps; ++i) {
1042 const MCOperand &MO = MI.getOperand(i);
1043 if (!MO.isReg()) continue;
1044 unsigned Reg = MO.getReg();
1045 if (!X86II::isX86_64NonExtLowByteReg(Reg)) continue;
1046 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
1047 // that returns non-zero.
1048 REX |= 0x40; // REX fixed encoding prefix
1052 switch (TSFlags & X86II::FormMask) {
1053 case X86II::MRMSrcReg:
1054 if (MI.getOperand(0).isReg() &&
1055 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
1056 REX |= 1 << 2; // set REX.R
1057 i = isTwoAddr ? 2 : 1;
1058 for (; i != NumOps; ++i) {
1059 const MCOperand &MO = MI.getOperand(i);
1060 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
1061 REX |= 1 << 0; // set REX.B
1064 case X86II::MRMSrcMem: {
1065 if (MI.getOperand(0).isReg() &&
1066 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
1067 REX |= 1 << 2; // set REX.R
1069 i = isTwoAddr ? 2 : 1;
1070 for (; i != NumOps; ++i) {
1071 const MCOperand &MO = MI.getOperand(i);
1073 if (X86II::isX86_64ExtendedReg(MO.getReg()))
1074 REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1)
1080 case X86II::MRM0m: case X86II::MRM1m:
1081 case X86II::MRM2m: case X86II::MRM3m:
1082 case X86II::MRM4m: case X86II::MRM5m:
1083 case X86II::MRM6m: case X86II::MRM7m:
1084 case X86II::MRMDestMem: {
1085 unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands);
1086 i = isTwoAddr ? 1 : 0;
1087 if (NumOps > e && MI.getOperand(e).isReg() &&
1088 X86II::isX86_64ExtendedReg(MI.getOperand(e).getReg()))
1089 REX |= 1 << 2; // set REX.R
1091 for (; i != e; ++i) {
1092 const MCOperand &MO = MI.getOperand(i);
1094 if (X86II::isX86_64ExtendedReg(MO.getReg()))
1095 REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1)
1102 if (MI.getOperand(0).isReg() &&
1103 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
1104 REX |= 1 << 0; // set REX.B
1105 i = isTwoAddr ? 2 : 1;
1106 for (unsigned e = NumOps; i != e; ++i) {
1107 const MCOperand &MO = MI.getOperand(i);
1108 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
1109 REX |= 1 << 2; // set REX.R
1116 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
1117 void X86MCCodeEmitter::EmitSegmentOverridePrefix(uint64_t TSFlags,
1118 unsigned &CurByte, int MemOperand,
1120 raw_ostream &OS) const {
1121 switch (TSFlags & X86II::SegOvrMask) {
1122 default: llvm_unreachable("Invalid segment!");
1124 // No segment override, check for explicit one on memory operand.
1125 if (MemOperand != -1) { // If the instruction has a memory operand.
1126 switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
1127 default: llvm_unreachable("Unknown segment register!");
1129 case X86::CS: EmitByte(0x2E, CurByte, OS); break;
1130 case X86::SS: EmitByte(0x36, CurByte, OS); break;
1131 case X86::DS: EmitByte(0x3E, CurByte, OS); break;
1132 case X86::ES: EmitByte(0x26, CurByte, OS); break;
1133 case X86::FS: EmitByte(0x64, CurByte, OS); break;
1134 case X86::GS: EmitByte(0x65, CurByte, OS); break;
1139 EmitByte(0x64, CurByte, OS);
1142 EmitByte(0x65, CurByte, OS);
1147 /// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode.
1149 /// MemOperand is the operand # of the start of a memory operand if present. If
1150 /// Not present, it is -1.
1151 void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
1152 int MemOperand, const MCInst &MI,
1153 const MCInstrDesc &Desc,
1154 raw_ostream &OS) const {
1156 // Emit the lock opcode prefix as needed.
1157 if (TSFlags & X86II::LOCK)
1158 EmitByte(0xF0, CurByte, OS);
1160 // Emit segment override opcode prefix as needed.
1161 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
1163 // Emit the repeat opcode prefix as needed.
1164 if ((TSFlags & X86II::Op0Mask) == X86II::REP)
1165 EmitByte(0xF3, CurByte, OS);
1167 // Emit the address size opcode prefix as needed.
1168 bool need_address_override;
1169 if (TSFlags & X86II::AdSize) {
1170 need_address_override = true;
1171 } else if (MemOperand == -1) {
1172 need_address_override = false;
1173 } else if (is64BitMode()) {
1174 assert(!Is16BitMemOperand(MI, MemOperand));
1175 need_address_override = Is32BitMemOperand(MI, MemOperand);
1176 } else if (is32BitMode()) {
1177 assert(!Is64BitMemOperand(MI, MemOperand));
1178 need_address_override = Is16BitMemOperand(MI, MemOperand);
1180 need_address_override = false;
1183 if (need_address_override)
1184 EmitByte(0x67, CurByte, OS);
1186 // Emit the operand size opcode prefix as needed.
1187 if (TSFlags & X86II::OpSize)
1188 EmitByte(0x66, CurByte, OS);
1190 bool Need0FPrefix = false;
1191 switch (TSFlags & X86II::Op0Mask) {
1192 default: llvm_unreachable("Invalid prefix!");
1193 case 0: break; // No prefix!
1194 case X86II::REP: break; // already handled.
1195 case X86II::TB: // Two-byte opcode prefix
1196 case X86II::T8: // 0F 38
1197 case X86II::TA: // 0F 3A
1198 case X86II::A6: // 0F A6
1199 case X86II::A7: // 0F A7
1200 Need0FPrefix = true;
1202 case X86II::XS: // F3 0F
1203 case X86II::T8XS: // F3 0F 38
1204 EmitByte(0xF3, CurByte, OS);
1205 Need0FPrefix = true;
1207 case X86II::XD: // F2 0F
1208 case X86II::T8XD: // F2 0F 38
1209 case X86II::TAXD: // F2 0F 3A
1210 EmitByte(0xF2, CurByte, OS);
1211 Need0FPrefix = true;
1221 EmitByte(0xD8+(((TSFlags & X86II::Op0Mask) - X86II::D8) >> X86II::Op0Shift),
1226 // Handle REX prefix.
1227 // FIXME: Can this come before F2 etc to simplify emission?
1228 if (is64BitMode()) {
1229 if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc))
1230 EmitByte(0x40 | REX, CurByte, OS);
1233 // 0x0F escape code must be emitted just before the opcode.
1235 EmitByte(0x0F, CurByte, OS);
1237 // FIXME: Pull this up into previous switch if REX can be moved earlier.
1238 switch (TSFlags & X86II::Op0Mask) {
1239 case X86II::T8XS: // F3 0F 38
1240 case X86II::T8XD: // F2 0F 38
1241 case X86II::T8: // 0F 38
1242 EmitByte(0x38, CurByte, OS);
1244 case X86II::TAXD: // F2 0F 3A
1245 case X86II::TA: // 0F 3A
1246 EmitByte(0x3A, CurByte, OS);
1248 case X86II::A6: // 0F A6
1249 EmitByte(0xA6, CurByte, OS);
1251 case X86II::A7: // 0F A7
1252 EmitByte(0xA7, CurByte, OS);
1257 void X86MCCodeEmitter::
1258 EncodeInstruction(const MCInst &MI, raw_ostream &OS,
1259 SmallVectorImpl<MCFixup> &Fixups) const {
1260 unsigned Opcode = MI.getOpcode();
1261 const MCInstrDesc &Desc = MCII.get(Opcode);
1262 uint64_t TSFlags = Desc.TSFlags;
1264 // Pseudo instructions don't get encoded.
1265 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
1268 unsigned NumOps = Desc.getNumOperands();
1269 unsigned CurOp = X86II::getOperandBias(Desc);
1271 // Keep track of the current byte being emitted.
1272 unsigned CurByte = 0;
1274 // Is this instruction encoded using the AVX VEX prefix?
1275 bool HasVEXPrefix = (TSFlags >> X86II::VEXShift) & X86II::VEX;
1277 // It uses the VEX.VVVV field?
1278 bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
1279 bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
1280 bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4;
1281 const unsigned MemOp4_I8IMMOperand = 2;
1283 // It uses the EVEX.aaa field?
1284 bool HasEVEX = (TSFlags >> X86II::VEXShift) & X86II::EVEX;
1285 bool HasEVEX_K = HasEVEX && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_K);
1286 bool HasEVEX_B = HasEVEX && ((TSFlags >> X86II::VEXShift) & X86II::EVEX_B);
1288 // Determine where the memory operand starts, if present.
1289 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags, Opcode);
1290 if (MemoryOperand != -1) MemoryOperand += CurOp;
1293 EmitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
1295 EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
1297 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
1299 if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
1300 BaseOpcode = 0x0F; // Weird 3DNow! encoding.
1302 unsigned SrcRegNum = 0;
1303 switch (TSFlags & X86II::FormMask) {
1304 default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
1305 llvm_unreachable("Unknown FormMask value in X86MCCodeEmitter!");
1307 llvm_unreachable("Pseudo instruction shouldn't be emitted");
1309 EmitByte(BaseOpcode, CurByte, OS);
1311 case X86II::RawFrmImm8:
1312 EmitByte(BaseOpcode, CurByte, OS);
1313 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1314 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1315 CurByte, OS, Fixups);
1316 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 1, FK_Data_1, CurByte,
1319 case X86II::RawFrmImm16:
1320 EmitByte(BaseOpcode, CurByte, OS);
1321 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1322 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
1323 CurByte, OS, Fixups);
1324 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(), 2, FK_Data_2, CurByte,
1328 case X86II::AddRegFrm:
1329 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
1332 case X86II::MRMDestReg:
1333 EmitByte(BaseOpcode, CurByte, OS);
1334 SrcRegNum = CurOp + 1;
1336 if (HasEVEX_K) // Skip writemask
1339 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1342 EmitRegModRMByte(MI.getOperand(CurOp),
1343 GetX86RegNum(MI.getOperand(SrcRegNum)), CurByte, OS);
1344 CurOp = SrcRegNum + 1;
1347 case X86II::MRMDestMem:
1348 EmitByte(BaseOpcode, CurByte, OS);
1349 SrcRegNum = CurOp + X86::AddrNumOperands;
1351 if (HasEVEX_K) // Skip writemask
1354 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1357 EmitMemModRMByte(MI, CurOp,
1358 GetX86RegNum(MI.getOperand(SrcRegNum)),
1359 TSFlags, CurByte, OS, Fixups);
1360 CurOp = SrcRegNum + 1;
1363 case X86II::MRMSrcReg:
1364 EmitByte(BaseOpcode, CurByte, OS);
1365 SrcRegNum = CurOp + 1;
1367 if (HasEVEX_K) // Skip writemask
1370 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1373 if (HasMemOp4) // Skip 2nd src (which is encoded in I8IMM)
1376 EmitRegModRMByte(MI.getOperand(SrcRegNum),
1377 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
1379 // 2 operands skipped with HasMemOp4, compensate accordingly
1380 CurOp = HasMemOp4 ? SrcRegNum : SrcRegNum + 1;
1383 // do not count the rounding control operand
1388 case X86II::MRMSrcMem: {
1389 int AddrOperands = X86::AddrNumOperands;
1390 unsigned FirstMemOp = CurOp+1;
1392 if (HasEVEX_K) { // Skip writemask
1399 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
1401 if (HasMemOp4) // Skip second register source (encoded in I8IMM)
1404 EmitByte(BaseOpcode, CurByte, OS);
1406 EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1407 TSFlags, CurByte, OS, Fixups);
1408 CurOp += AddrOperands + 1;
1414 case X86II::MRM0r: case X86II::MRM1r:
1415 case X86II::MRM2r: case X86II::MRM3r:
1416 case X86II::MRM4r: case X86II::MRM5r:
1417 case X86II::MRM6r: case X86II::MRM7r:
1418 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1420 EmitByte(BaseOpcode, CurByte, OS);
1421 EmitRegModRMByte(MI.getOperand(CurOp++),
1422 (TSFlags & X86II::FormMask)-X86II::MRM0r,
1425 case X86II::MRM0m: case X86II::MRM1m:
1426 case X86II::MRM2m: case X86II::MRM3m:
1427 case X86II::MRM4m: case X86II::MRM5m:
1428 case X86II::MRM6m: case X86II::MRM7m:
1429 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1431 EmitByte(BaseOpcode, CurByte, OS);
1432 EmitMemModRMByte(MI, CurOp, (TSFlags & X86II::FormMask)-X86II::MRM0m,
1433 TSFlags, CurByte, OS, Fixups);
1434 CurOp += X86::AddrNumOperands;
1436 case X86II::MRM_C1: case X86II::MRM_C2: case X86II::MRM_C3:
1437 case X86II::MRM_C4: case X86II::MRM_C8: case X86II::MRM_C9:
1438 case X86II::MRM_CA: case X86II::MRM_CB: case X86II::MRM_D0:
1439 case X86II::MRM_D1: case X86II::MRM_D4: case X86II::MRM_D5:
1440 case X86II::MRM_D6: case X86II::MRM_D8: case X86II::MRM_D9:
1441 case X86II::MRM_DA: case X86II::MRM_DB: case X86II::MRM_DC:
1442 case X86II::MRM_DD: case X86II::MRM_DE: case X86II::MRM_DF:
1443 case X86II::MRM_E8: case X86II::MRM_F0: case X86II::MRM_F8:
1445 EmitByte(BaseOpcode, CurByte, OS);
1448 switch (TSFlags & X86II::FormMask) {
1449 default: llvm_unreachable("Invalid Form");
1450 case X86II::MRM_C1: MRM = 0xC1; break;
1451 case X86II::MRM_C2: MRM = 0xC2; break;
1452 case X86II::MRM_C3: MRM = 0xC3; break;
1453 case X86II::MRM_C4: MRM = 0xC4; break;
1454 case X86II::MRM_C8: MRM = 0xC8; break;
1455 case X86II::MRM_C9: MRM = 0xC9; break;
1456 case X86II::MRM_CA: MRM = 0xCA; break;
1457 case X86II::MRM_CB: MRM = 0xCB; break;
1458 case X86II::MRM_D0: MRM = 0xD0; break;
1459 case X86II::MRM_D1: MRM = 0xD1; break;
1460 case X86II::MRM_D4: MRM = 0xD4; break;
1461 case X86II::MRM_D5: MRM = 0xD5; break;
1462 case X86II::MRM_D6: MRM = 0xD6; break;
1463 case X86II::MRM_D8: MRM = 0xD8; break;
1464 case X86II::MRM_D9: MRM = 0xD9; break;
1465 case X86II::MRM_DA: MRM = 0xDA; break;
1466 case X86II::MRM_DB: MRM = 0xDB; break;
1467 case X86II::MRM_DC: MRM = 0xDC; break;
1468 case X86II::MRM_DD: MRM = 0xDD; break;
1469 case X86II::MRM_DE: MRM = 0xDE; break;
1470 case X86II::MRM_DF: MRM = 0xDF; break;
1471 case X86II::MRM_E8: MRM = 0xE8; break;
1472 case X86II::MRM_F0: MRM = 0xF0; break;
1473 case X86II::MRM_F8: MRM = 0xF8; break;
1474 case X86II::MRM_F9: MRM = 0xF9; break;
1476 EmitByte(MRM, CurByte, OS);
1480 // If there is a remaining operand, it must be a trailing immediate. Emit it
1481 // according to the right size for the instruction. Some instructions
1482 // (SSE4a extrq and insertq) have two trailing immediates.
1483 while (CurOp != NumOps && NumOps - CurOp <= 2) {
1484 // The last source register of a 4 operand instruction in AVX is encoded
1485 // in bits[7:4] of a immediate byte.
1486 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) {
1487 const MCOperand &MO = MI.getOperand(HasMemOp4 ? MemOp4_I8IMMOperand
1490 unsigned RegNum = GetX86RegNum(MO) << 4;
1491 if (X86II::isX86_64ExtendedReg(MO.getReg()))
1493 // If there is an additional 5th operand it must be an immediate, which
1494 // is encoded in bits[3:0]
1495 if (CurOp != NumOps) {
1496 const MCOperand &MIMM = MI.getOperand(CurOp++);
1498 unsigned Val = MIMM.getImm();
1499 assert(Val < 16 && "Immediate operand value out of range");
1503 EmitImmediate(MCOperand::CreateImm(RegNum), MI.getLoc(), 1, FK_Data_1,
1504 CurByte, OS, Fixups);
1507 // FIXME: Is there a better way to know that we need a signed relocation?
1508 if (MI.getOpcode() == X86::ADD64ri32 ||
1509 MI.getOpcode() == X86::MOV64ri32 ||
1510 MI.getOpcode() == X86::MOV64mi32 ||
1511 MI.getOpcode() == X86::PUSH64i32)
1512 FixupKind = X86::reloc_signed_4byte;
1514 FixupKind = getImmFixupKind(TSFlags);
1515 EmitImmediate(MI.getOperand(CurOp++), MI.getLoc(),
1516 X86II::getSizeOfImm(TSFlags), MCFixupKind(FixupKind),
1517 CurByte, OS, Fixups);
1521 if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
1522 EmitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS);
1526 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
1527 errs() << "Cannot encode all operands of: ";