1 //===-- X86/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/MCExpr.h"
20 #include "llvm/MC/MCInst.h"
21 #include "llvm/MC/MCInstrInfo.h"
22 #include "llvm/MC/MCRegisterInfo.h"
23 #include "llvm/MC/MCSubtargetInfo.h"
24 #include "llvm/MC/MCSymbol.h"
25 #include "llvm/Support/raw_ostream.h"
30 class X86MCCodeEmitter : public MCCodeEmitter {
31 X86MCCodeEmitter(const X86MCCodeEmitter &); // DO NOT IMPLEMENT
32 void operator=(const X86MCCodeEmitter &); // DO NOT IMPLEMENT
33 const MCInstrInfo &MCII;
34 const MCSubtargetInfo &STI;
37 X86MCCodeEmitter(const MCInstrInfo &mcii, const MCSubtargetInfo &sti,
39 : MCII(mcii), STI(sti), Ctx(ctx) {
42 ~X86MCCodeEmitter() {}
44 bool is64BitMode() const {
45 // FIXME: Can tablegen auto-generate this?
46 return (STI.getFeatureBits() & X86::Mode64Bit) != 0;
49 static unsigned GetX86RegNum(const MCOperand &MO) {
50 return X86_MC::getX86RegNum(MO.getReg());
53 // On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the range
54 // 0-7 and the difference between the 2 groups is given by the REX prefix.
55 // In the VEX prefix, registers are seen sequencially from 0-15 and encoded
56 // in 1's complement form, example:
58 // ModRM field => XMM9 => 1
59 // VEX.VVVV => XMM9 => ~9
61 // See table 4-35 of Intel AVX Programming Reference for details.
62 static unsigned char getVEXRegisterEncoding(const MCInst &MI,
64 unsigned SrcReg = MI.getOperand(OpNum).getReg();
65 unsigned SrcRegNum = GetX86RegNum(MI.getOperand(OpNum));
66 if (X86II::isX86_64ExtendedReg(SrcReg))
69 // The registers represented through VEX_VVVV should
70 // be encoded in 1's complement form.
71 return (~SrcRegNum) & 0xf;
74 void EmitByte(unsigned char C, unsigned &CurByte, raw_ostream &OS) const {
79 void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte,
80 raw_ostream &OS) const {
81 // Output the constant in little endian byte order.
82 for (unsigned i = 0; i != Size; ++i) {
83 EmitByte(Val & 255, CurByte, OS);
88 void EmitImmediate(const MCOperand &Disp,
89 unsigned ImmSize, MCFixupKind FixupKind,
90 unsigned &CurByte, raw_ostream &OS,
91 SmallVectorImpl<MCFixup> &Fixups,
92 int ImmOffset = 0) const;
94 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
96 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
97 return RM | (RegOpcode << 3) | (Mod << 6);
100 void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld,
101 unsigned &CurByte, raw_ostream &OS) const {
102 EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS);
105 void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base,
106 unsigned &CurByte, raw_ostream &OS) const {
107 // SIB byte is in the same format as the ModRMByte.
108 EmitByte(ModRMByte(SS, Index, Base), CurByte, OS);
112 void EmitMemModRMByte(const MCInst &MI, unsigned Op,
113 unsigned RegOpcodeField,
114 uint64_t TSFlags, unsigned &CurByte, raw_ostream &OS,
115 SmallVectorImpl<MCFixup> &Fixups) const;
117 void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
118 SmallVectorImpl<MCFixup> &Fixups) const;
120 void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
121 const MCInst &MI, const MCInstrDesc &Desc,
122 raw_ostream &OS) const;
124 void EmitSegmentOverridePrefix(uint64_t TSFlags, unsigned &CurByte,
125 int MemOperand, const MCInst &MI,
126 raw_ostream &OS) const;
128 void EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand,
129 const MCInst &MI, const MCInstrDesc &Desc,
130 raw_ostream &OS) const;
133 } // end anonymous namespace
136 MCCodeEmitter *llvm::createX86MCCodeEmitter(const MCInstrInfo &MCII,
137 const MCSubtargetInfo &STI,
139 return new X86MCCodeEmitter(MCII, STI, Ctx);
142 /// isDisp8 - Return true if this signed displacement fits in a 8-bit
143 /// sign-extended field.
144 static bool isDisp8(int Value) {
145 return Value == (signed char)Value;
148 /// getImmFixupKind - Return the appropriate fixup kind to use for an immediate
149 /// in an instruction with the specified TSFlags.
150 static MCFixupKind getImmFixupKind(uint64_t TSFlags) {
151 unsigned Size = X86II::getSizeOfImm(TSFlags);
152 bool isPCRel = X86II::isImmPCRel(TSFlags);
154 return MCFixup::getKindForSize(Size, isPCRel);
157 /// Is32BitMemOperand - Return true if the specified instruction with a memory
158 /// operand should emit the 0x67 prefix byte in 64-bit mode due to a 32-bit
159 /// memory operand. Op specifies the operand # of the memoperand.
160 static bool Is32BitMemOperand(const MCInst &MI, unsigned Op) {
161 const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
162 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
164 if ((BaseReg.getReg() != 0 &&
165 X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) ||
166 (IndexReg.getReg() != 0 &&
167 X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg())))
172 /// StartsWithGlobalOffsetTable - Return true for the simple cases where this
173 /// expression starts with _GLOBAL_OFFSET_TABLE_. This is a needed to support
174 /// PIC on ELF i386 as that symbol is magic. We check only simple case that
175 /// are know to be used: _GLOBAL_OFFSET_TABLE_ by itself or at the start
176 /// of a binary expression.
177 static bool StartsWithGlobalOffsetTable(const MCExpr *Expr) {
178 if (Expr->getKind() == MCExpr::Binary) {
179 const MCBinaryExpr *BE = static_cast<const MCBinaryExpr *>(Expr);
183 if (Expr->getKind() != MCExpr::SymbolRef)
186 const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr);
187 const MCSymbol &S = Ref->getSymbol();
188 return S.getName() == "_GLOBAL_OFFSET_TABLE_";
191 void X86MCCodeEmitter::
192 EmitImmediate(const MCOperand &DispOp, unsigned Size, MCFixupKind FixupKind,
193 unsigned &CurByte, raw_ostream &OS,
194 SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const {
195 const MCExpr *Expr = NULL;
196 if (DispOp.isImm()) {
197 // If this is a simple integer displacement that doesn't require a
198 // relocation, emit it now.
199 if (FixupKind != FK_PCRel_1 &&
200 FixupKind != FK_PCRel_2 &&
201 FixupKind != FK_PCRel_4) {
202 EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS);
205 Expr = MCConstantExpr::Create(DispOp.getImm(), Ctx);
207 Expr = DispOp.getExpr();
210 // If we have an immoffset, add it to the expression.
211 if ((FixupKind == FK_Data_4 ||
212 FixupKind == MCFixupKind(X86::reloc_signed_4byte)) &&
213 StartsWithGlobalOffsetTable(Expr)) {
214 assert(ImmOffset == 0);
216 FixupKind = MCFixupKind(X86::reloc_global_offset_table);
220 // If the fixup is pc-relative, we need to bias the value to be relative to
221 // the start of the field, not the end of the field.
222 if (FixupKind == FK_PCRel_4 ||
223 FixupKind == MCFixupKind(X86::reloc_riprel_4byte) ||
224 FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load))
226 if (FixupKind == FK_PCRel_2)
228 if (FixupKind == FK_PCRel_1)
232 Expr = MCBinaryExpr::CreateAdd(Expr, MCConstantExpr::Create(ImmOffset, Ctx),
235 // Emit a symbolic constant as a fixup and 4 zeros.
236 Fixups.push_back(MCFixup::Create(CurByte, Expr, FixupKind));
237 EmitConstant(0, Size, CurByte, OS);
240 void X86MCCodeEmitter::EmitMemModRMByte(const MCInst &MI, unsigned Op,
241 unsigned RegOpcodeField,
242 uint64_t TSFlags, unsigned &CurByte,
244 SmallVectorImpl<MCFixup> &Fixups) const{
245 const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp);
246 const MCOperand &Base = MI.getOperand(Op+X86::AddrBaseReg);
247 const MCOperand &Scale = MI.getOperand(Op+X86::AddrScaleAmt);
248 const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
249 unsigned BaseReg = Base.getReg();
251 // Handle %rip relative addressing.
252 if (BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
253 assert(is64BitMode() && "Rip-relative addressing requires 64-bit mode");
254 assert(IndexReg.getReg() == 0 && "Invalid rip-relative address");
255 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
257 unsigned FixupKind = X86::reloc_riprel_4byte;
259 // movq loads are handled with a special relocation form which allows the
260 // linker to eliminate some loads for GOT references which end up in the
261 // same linkage unit.
262 if (MI.getOpcode() == X86::MOV64rm)
263 FixupKind = X86::reloc_riprel_4byte_movq_load;
265 // rip-relative addressing is actually relative to the *next* instruction.
266 // Since an immediate can follow the mod/rm byte for an instruction, this
267 // means that we need to bias the immediate field of the instruction with
268 // the size of the immediate field. If we have this case, add it into the
269 // expression to emit.
270 int ImmSize = X86II::hasImm(TSFlags) ? X86II::getSizeOfImm(TSFlags) : 0;
272 EmitImmediate(Disp, 4, MCFixupKind(FixupKind),
273 CurByte, OS, Fixups, -ImmSize);
277 unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U;
279 // Determine whether a SIB byte is needed.
280 // If no BaseReg, issue a RIP relative instruction only if the MCE can
281 // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
282 // 2-7) and absolute references.
284 if (// The SIB byte must be used if there is an index register.
285 IndexReg.getReg() == 0 &&
286 // The SIB byte must be used if the base is ESP/RSP/R12, all of which
287 // encode to an R/M value of 4, which indicates that a SIB byte is
289 BaseRegNo != N86::ESP &&
290 // If there is no base register and we're in 64-bit mode, we need a SIB
291 // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
292 (!is64BitMode() || BaseReg != 0)) {
294 if (BaseReg == 0) { // [disp32] in X86-32 mode
295 EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS);
296 EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups);
300 // If the base is not EBP/ESP and there is no displacement, use simple
301 // indirect register encoding, this handles addresses like [EAX]. The
302 // encoding for [EBP] with no displacement means [disp32] so we handle it
303 // by emitting a displacement of 0 below.
304 if (Disp.isImm() && Disp.getImm() == 0 && BaseRegNo != N86::EBP) {
305 EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS);
309 // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
310 if (Disp.isImm() && isDisp8(Disp.getImm())) {
311 EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS);
312 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
316 // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
317 EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS);
318 EmitImmediate(Disp, 4, MCFixupKind(X86::reloc_signed_4byte), CurByte, OS,
323 // We need a SIB byte, so start by outputting the ModR/M byte first
324 assert(IndexReg.getReg() != X86::ESP &&
325 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
327 bool ForceDisp32 = false;
328 bool ForceDisp8 = false;
330 // If there is no base register, we emit the special case SIB byte with
331 // MOD=0, BASE=5, to JUST get the index, scale, and displacement.
332 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
334 } else if (!Disp.isImm()) {
335 // Emit the normal disp32 encoding.
336 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
338 } else if (Disp.getImm() == 0 &&
339 // Base reg can't be anything that ends up with '5' as the base
340 // reg, it is the magic [*] nomenclature that indicates no base.
341 BaseRegNo != N86::EBP) {
342 // Emit no displacement ModR/M byte
343 EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS);
344 } else if (isDisp8(Disp.getImm())) {
345 // Emit the disp8 encoding.
346 EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS);
347 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
349 // Emit the normal disp32 encoding.
350 EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS);
353 // Calculate what the SS field value should be...
354 static const unsigned SSTable[] = { ~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3 };
355 unsigned SS = SSTable[Scale.getImm()];
358 // Handle the SIB byte for the case where there is no base, see Intel
359 // Manual 2A, table 2-7. The displacement has already been output.
361 if (IndexReg.getReg())
362 IndexRegNo = GetX86RegNum(IndexReg);
363 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
365 EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS);
368 if (IndexReg.getReg())
369 IndexRegNo = GetX86RegNum(IndexReg);
371 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
372 EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS);
375 // Do we need to output a displacement?
377 EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups);
378 else if (ForceDisp32 || Disp.getImm() != 0)
379 EmitImmediate(Disp, 4, MCFixupKind(X86::reloc_signed_4byte), CurByte, OS,
383 /// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix
385 void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
386 int MemOperand, const MCInst &MI,
387 const MCInstrDesc &Desc,
388 raw_ostream &OS) const {
389 bool HasVEX_4V = false;
390 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_4V)
393 // VEX_R: opcode externsion equivalent to REX.R in
394 // 1's complement (inverted) form
396 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
397 // 0: Same as REX_R=1 (64 bit mode only)
399 unsigned char VEX_R = 0x1;
401 // VEX_X: equivalent to REX.X, only used when a
402 // register is used for index in SIB Byte.
404 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
405 // 0: Same as REX.X=1 (64-bit mode only)
406 unsigned char VEX_X = 0x1;
410 // 1: Same as REX_B=0 (ignored in 32-bit mode)
411 // 0: Same as REX_B=1 (64 bit mode only)
413 unsigned char VEX_B = 0x1;
415 // VEX_W: opcode specific (use like REX.W, or used for
416 // opcode extension, or ignored, depending on the opcode byte)
417 unsigned char VEX_W = 0;
419 // VEX_5M (VEX m-mmmmm field):
421 // 0b00000: Reserved for future use
422 // 0b00001: implied 0F leading opcode
423 // 0b00010: implied 0F 38 leading opcode bytes
424 // 0b00011: implied 0F 3A leading opcode bytes
425 // 0b00100-0b11111: Reserved for future use
427 unsigned char VEX_5M = 0x1;
429 // VEX_4V (VEX vvvv field): a register specifier
430 // (in 1's complement form) or 1111 if unused.
431 unsigned char VEX_4V = 0xf;
433 // VEX_L (Vector Length):
435 // 0: scalar or 128-bit vector
438 unsigned char VEX_L = 0;
440 // VEX_PP: opcode extension providing equivalent
441 // functionality of a SIMD prefix
448 unsigned char VEX_PP = 0;
450 // FIXME: BEXTR uses VEX.vvvv for Operand 3 instead of Operand 2
451 unsigned Opcode = MI.getOpcode();
452 bool IsBEXTR = (Opcode == X86::BEXTR32rr || Opcode == X86::BEXTR32rm ||
453 Opcode == X86::BEXTR64rr || Opcode == X86::BEXTR64rm);
455 // Encode the operand size opcode prefix as needed.
456 if (TSFlags & X86II::OpSize)
459 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_W)
462 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_L)
465 switch (TSFlags & X86II::Op0Mask) {
466 default: assert(0 && "Invalid prefix!");
467 case X86II::T8: // 0F 38
470 case X86II::TA: // 0F 3A
473 case X86II::TF: // F2 0F 38
477 case X86II::XS: // F3 0F
480 case X86II::XD: // F2 0F
483 case X86II::A6: // Bypass: Not used by VEX
484 case X86II::A7: // Bypass: Not used by VEX
485 case X86II::TB: // Bypass: Not used by VEX
490 // Set the vector length to 256-bit if YMM0-YMM15 is used
491 for (unsigned i = 0; i != MI.getNumOperands(); ++i) {
492 if (!MI.getOperand(i).isReg())
494 unsigned SrcReg = MI.getOperand(i).getReg();
495 if (SrcReg >= X86::YMM0 && SrcReg <= X86::YMM15)
499 // Classify VEX_B, VEX_4V, VEX_R, VEX_X
501 switch (TSFlags & X86II::FormMask) {
502 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
503 case X86II::MRMDestMem: {
504 // MRMDestMem instructions forms:
505 // MemAddr, src1(ModR/M)
506 // MemAddr, src1(VEX_4V), src2(ModR/M)
507 // MemAddr, src1(ModR/M), imm8
509 if (X86II::isX86_64ExtendedReg(MI.getOperand(X86::AddrBaseReg).getReg()))
511 if (X86II::isX86_64ExtendedReg(MI.getOperand(X86::AddrIndexReg).getReg()))
514 CurOp = X86::AddrNumOperands;
516 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
518 const MCOperand &MO = MI.getOperand(CurOp);
519 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
523 case X86II::MRMSrcMem:
524 // MRMSrcMem instructions forms:
525 // src1(ModR/M), MemAddr
526 // src1(ModR/M), src2(VEX_4V), MemAddr
527 // src1(ModR/M), MemAddr, imm8
528 // src1(ModR/M), MemAddr, src2(VEX_I8IMM)
530 if (X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
533 // FIXME: BEXTR uses VEX.vvvv for Operand 3
534 if (HasVEX_4V && !IsBEXTR)
535 VEX_4V = getVEXRegisterEncoding(MI, 1);
537 if (X86II::isX86_64ExtendedReg(
538 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
540 if (X86II::isX86_64ExtendedReg(
541 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
545 VEX_4V = getVEXRegisterEncoding(MI, X86::AddrNumOperands+1);
547 case X86II::MRM0m: case X86II::MRM1m:
548 case X86II::MRM2m: case X86II::MRM3m:
549 case X86II::MRM4m: case X86II::MRM5m:
550 case X86II::MRM6m: case X86II::MRM7m: {
551 // MRM[0-9]m instructions forms:
553 // src1(VEX_4V), MemAddr
555 VEX_4V = getVEXRegisterEncoding(MI, 0);
557 if (X86II::isX86_64ExtendedReg(
558 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
560 if (X86II::isX86_64ExtendedReg(
561 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
565 case X86II::MRMSrcReg:
566 // MRMSrcReg instructions forms:
567 // dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
568 // dst(ModR/M), src1(ModR/M)
569 // dst(ModR/M), src1(ModR/M), imm8
571 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
575 // FIXME: BEXTR uses VEX.vvvv for Operand 3
576 if (HasVEX_4V && !IsBEXTR)
577 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
578 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
582 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
584 case X86II::MRMDestReg:
585 // MRMDestReg instructions forms:
586 // dst(ModR/M), src(ModR/M)
587 // dst(ModR/M), src(ModR/M), imm8
588 if (X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
590 if (X86II::isX86_64ExtendedReg(MI.getOperand(1).getReg()))
593 case X86II::MRM0r: case X86II::MRM1r:
594 case X86II::MRM2r: case X86II::MRM3r:
595 case X86II::MRM4r: case X86II::MRM5r:
596 case X86II::MRM6r: case X86II::MRM7r:
597 // MRM0r-MRM7r instructions forms:
598 // dst(VEX_4V), src(ModR/M), imm8
599 VEX_4V = getVEXRegisterEncoding(MI, 0);
600 if (X86II::isX86_64ExtendedReg(MI.getOperand(1).getReg()))
607 // Emit segment override opcode prefix as needed.
608 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
610 // VEX opcode prefix can have 2 or 3 bytes
613 // +-----+ +--------------+ +-------------------+
614 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
615 // +-----+ +--------------+ +-------------------+
617 // +-----+ +-------------------+
618 // | C5h | | R | vvvv | L | pp |
619 // +-----+ +-------------------+
621 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
623 if (VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) { // 2 byte VEX prefix
624 EmitByte(0xC5, CurByte, OS);
625 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
630 EmitByte(0xC4, CurByte, OS);
631 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
632 EmitByte(LastByte | (VEX_W << 7), CurByte, OS);
635 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
636 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
637 /// size, and 3) use of X86-64 extended registers.
638 static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
639 const MCInstrDesc &Desc) {
641 if (TSFlags & X86II::REX_W)
642 REX |= 1 << 3; // set REX.W
644 if (MI.getNumOperands() == 0) return REX;
646 unsigned NumOps = MI.getNumOperands();
647 // FIXME: MCInst should explicitize the two-addrness.
648 bool isTwoAddr = NumOps > 1 &&
649 Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1;
651 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
652 unsigned i = isTwoAddr ? 1 : 0;
653 for (; i != NumOps; ++i) {
654 const MCOperand &MO = MI.getOperand(i);
655 if (!MO.isReg()) continue;
656 unsigned Reg = MO.getReg();
657 if (!X86II::isX86_64NonExtLowByteReg(Reg)) continue;
658 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
659 // that returns non-zero.
660 REX |= 0x40; // REX fixed encoding prefix
664 switch (TSFlags & X86II::FormMask) {
665 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
666 case X86II::MRMSrcReg:
667 if (MI.getOperand(0).isReg() &&
668 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
669 REX |= 1 << 2; // set REX.R
670 i = isTwoAddr ? 2 : 1;
671 for (; i != NumOps; ++i) {
672 const MCOperand &MO = MI.getOperand(i);
673 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
674 REX |= 1 << 0; // set REX.B
677 case X86II::MRMSrcMem: {
678 if (MI.getOperand(0).isReg() &&
679 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
680 REX |= 1 << 2; // set REX.R
682 i = isTwoAddr ? 2 : 1;
683 for (; i != NumOps; ++i) {
684 const MCOperand &MO = MI.getOperand(i);
686 if (X86II::isX86_64ExtendedReg(MO.getReg()))
687 REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1)
693 case X86II::MRM0m: case X86II::MRM1m:
694 case X86II::MRM2m: case X86II::MRM3m:
695 case X86II::MRM4m: case X86II::MRM5m:
696 case X86II::MRM6m: case X86II::MRM7m:
697 case X86II::MRMDestMem: {
698 unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands);
699 i = isTwoAddr ? 1 : 0;
700 if (NumOps > e && MI.getOperand(e).isReg() &&
701 X86II::isX86_64ExtendedReg(MI.getOperand(e).getReg()))
702 REX |= 1 << 2; // set REX.R
704 for (; i != e; ++i) {
705 const MCOperand &MO = MI.getOperand(i);
707 if (X86II::isX86_64ExtendedReg(MO.getReg()))
708 REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1)
715 if (MI.getOperand(0).isReg() &&
716 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
717 REX |= 1 << 0; // set REX.B
718 i = isTwoAddr ? 2 : 1;
719 for (unsigned e = NumOps; i != e; ++i) {
720 const MCOperand &MO = MI.getOperand(i);
721 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
722 REX |= 1 << 2; // set REX.R
729 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
730 void X86MCCodeEmitter::EmitSegmentOverridePrefix(uint64_t TSFlags,
731 unsigned &CurByte, int MemOperand,
733 raw_ostream &OS) const {
734 switch (TSFlags & X86II::SegOvrMask) {
735 default: assert(0 && "Invalid segment!");
737 // No segment override, check for explicit one on memory operand.
738 if (MemOperand != -1) { // If the instruction has a memory operand.
739 switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
740 default: assert(0 && "Unknown segment register!");
742 case X86::CS: EmitByte(0x2E, CurByte, OS); break;
743 case X86::SS: EmitByte(0x36, CurByte, OS); break;
744 case X86::DS: EmitByte(0x3E, CurByte, OS); break;
745 case X86::ES: EmitByte(0x26, CurByte, OS); break;
746 case X86::FS: EmitByte(0x64, CurByte, OS); break;
747 case X86::GS: EmitByte(0x65, CurByte, OS); break;
752 EmitByte(0x64, CurByte, OS);
755 EmitByte(0x65, CurByte, OS);
760 /// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode.
762 /// MemOperand is the operand # of the start of a memory operand if present. If
763 /// Not present, it is -1.
764 void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
765 int MemOperand, const MCInst &MI,
766 const MCInstrDesc &Desc,
767 raw_ostream &OS) const {
769 // Emit the lock opcode prefix as needed.
770 if (TSFlags & X86II::LOCK)
771 EmitByte(0xF0, CurByte, OS);
773 // Emit segment override opcode prefix as needed.
774 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
776 // Emit the repeat opcode prefix as needed.
777 if ((TSFlags & X86II::Op0Mask) == X86II::REP)
778 EmitByte(0xF3, CurByte, OS);
780 // Emit the address size opcode prefix as needed.
781 if ((TSFlags & X86II::AdSize) ||
782 (MemOperand != -1 && is64BitMode() && Is32BitMemOperand(MI, MemOperand)))
783 EmitByte(0x67, CurByte, OS);
785 // Emit the operand size opcode prefix as needed.
786 if (TSFlags & X86II::OpSize)
787 EmitByte(0x66, CurByte, OS);
789 bool Need0FPrefix = false;
790 switch (TSFlags & X86II::Op0Mask) {
791 default: assert(0 && "Invalid prefix!");
792 case 0: break; // No prefix!
793 case X86II::REP: break; // already handled.
794 case X86II::TB: // Two-byte opcode prefix
795 case X86II::T8: // 0F 38
796 case X86II::TA: // 0F 3A
797 case X86II::A6: // 0F A6
798 case X86II::A7: // 0F A7
801 case X86II::TF: // F2 0F 38
802 EmitByte(0xF2, CurByte, OS);
805 case X86II::XS: // F3 0F
806 EmitByte(0xF3, CurByte, OS);
809 case X86II::XD: // F2 0F
810 EmitByte(0xF2, CurByte, OS);
813 case X86II::D8: EmitByte(0xD8, CurByte, OS); break;
814 case X86II::D9: EmitByte(0xD9, CurByte, OS); break;
815 case X86II::DA: EmitByte(0xDA, CurByte, OS); break;
816 case X86II::DB: EmitByte(0xDB, CurByte, OS); break;
817 case X86II::DC: EmitByte(0xDC, CurByte, OS); break;
818 case X86II::DD: EmitByte(0xDD, CurByte, OS); break;
819 case X86II::DE: EmitByte(0xDE, CurByte, OS); break;
820 case X86II::DF: EmitByte(0xDF, CurByte, OS); break;
823 // Handle REX prefix.
824 // FIXME: Can this come before F2 etc to simplify emission?
826 if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc))
827 EmitByte(0x40 | REX, CurByte, OS);
830 // 0x0F escape code must be emitted just before the opcode.
832 EmitByte(0x0F, CurByte, OS);
834 // FIXME: Pull this up into previous switch if REX can be moved earlier.
835 switch (TSFlags & X86II::Op0Mask) {
836 case X86II::TF: // F2 0F 38
837 case X86II::T8: // 0F 38
838 EmitByte(0x38, CurByte, OS);
840 case X86II::TA: // 0F 3A
841 EmitByte(0x3A, CurByte, OS);
843 case X86II::A6: // 0F A6
844 EmitByte(0xA6, CurByte, OS);
846 case X86II::A7: // 0F A7
847 EmitByte(0xA7, CurByte, OS);
852 void X86MCCodeEmitter::
853 EncodeInstruction(const MCInst &MI, raw_ostream &OS,
854 SmallVectorImpl<MCFixup> &Fixups) const {
855 unsigned Opcode = MI.getOpcode();
856 const MCInstrDesc &Desc = MCII.get(Opcode);
857 uint64_t TSFlags = Desc.TSFlags;
859 // Pseudo instructions don't get encoded.
860 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
863 // If this is a two-address instruction, skip one of the register operands.
864 // FIXME: This should be handled during MCInst lowering.
865 unsigned NumOps = Desc.getNumOperands();
867 if (NumOps > 1 && Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1)
869 else if (NumOps > 2 && Desc.getOperandConstraint(NumOps-1, MCOI::TIED_TO)== 0)
870 // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32
873 // Keep track of the current byte being emitted.
874 unsigned CurByte = 0;
876 // Is this instruction encoded using the AVX VEX prefix?
877 bool HasVEXPrefix = false;
879 // It uses the VEX.VVVV field?
880 bool HasVEX_4V = false;
882 if ((TSFlags >> X86II::VEXShift) & X86II::VEX)
884 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_4V)
887 // Determine where the memory operand starts, if present.
888 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags, Opcode);
889 if (MemoryOperand != -1) MemoryOperand += CurOp;
892 EmitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
894 EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
896 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
898 if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
899 BaseOpcode = 0x0F; // Weird 3DNow! encoding.
901 // FIXME: BEXTR uses VEX.vvvv for Operand 3 instead of Operand 2
902 bool IsBEXTR = (Opcode == X86::BEXTR32rr || Opcode == X86::BEXTR32rm ||
903 Opcode == X86::BEXTR64rr || Opcode == X86::BEXTR64rm);
905 unsigned SrcRegNum = 0;
906 switch (TSFlags & X86II::FormMask) {
907 case X86II::MRMInitReg:
908 assert(0 && "FIXME: Remove this form when the JIT moves to MCCodeEmitter!");
909 default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
910 assert(0 && "Unknown FormMask value in X86MCCodeEmitter!");
912 assert(0 && "Pseudo instruction shouldn't be emitted");
914 EmitByte(BaseOpcode, CurByte, OS);
916 case X86II::RawFrmImm8:
917 EmitByte(BaseOpcode, CurByte, OS);
918 EmitImmediate(MI.getOperand(CurOp++),
919 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
920 CurByte, OS, Fixups);
921 EmitImmediate(MI.getOperand(CurOp++), 1, FK_Data_1, CurByte, OS, Fixups);
923 case X86II::RawFrmImm16:
924 EmitByte(BaseOpcode, CurByte, OS);
925 EmitImmediate(MI.getOperand(CurOp++),
926 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
927 CurByte, OS, Fixups);
928 EmitImmediate(MI.getOperand(CurOp++), 2, FK_Data_2, CurByte, OS, Fixups);
931 case X86II::AddRegFrm:
932 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
935 case X86II::MRMDestReg:
936 EmitByte(BaseOpcode, CurByte, OS);
937 EmitRegModRMByte(MI.getOperand(CurOp),
938 GetX86RegNum(MI.getOperand(CurOp+1)), CurByte, OS);
942 case X86II::MRMDestMem:
943 EmitByte(BaseOpcode, CurByte, OS);
944 SrcRegNum = CurOp + X86::AddrNumOperands;
946 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
949 EmitMemModRMByte(MI, CurOp,
950 GetX86RegNum(MI.getOperand(SrcRegNum)),
951 TSFlags, CurByte, OS, Fixups);
952 CurOp = SrcRegNum + 1;
955 case X86II::MRMSrcReg:
956 EmitByte(BaseOpcode, CurByte, OS);
957 SrcRegNum = CurOp + 1;
959 if (HasVEX_4V && !IsBEXTR) // Skip 1st src (which is encoded in VEX_VVVV)
962 EmitRegModRMByte(MI.getOperand(SrcRegNum),
963 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
964 CurOp = SrcRegNum + 1;
969 case X86II::MRMSrcMem: {
970 int AddrOperands = X86::AddrNumOperands;
971 unsigned FirstMemOp = CurOp+1;
972 if (HasVEX_4V && !IsBEXTR) {
974 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
977 EmitByte(BaseOpcode, CurByte, OS);
979 EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
980 TSFlags, CurByte, OS, Fixups);
981 CurOp += AddrOperands + 1;
987 case X86II::MRM0r: case X86II::MRM1r:
988 case X86II::MRM2r: case X86II::MRM3r:
989 case X86II::MRM4r: case X86II::MRM5r:
990 case X86II::MRM6r: case X86II::MRM7r:
991 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
993 EmitByte(BaseOpcode, CurByte, OS);
994 EmitRegModRMByte(MI.getOperand(CurOp++),
995 (TSFlags & X86II::FormMask)-X86II::MRM0r,
998 case X86II::MRM0m: case X86II::MRM1m:
999 case X86II::MRM2m: case X86II::MRM3m:
1000 case X86II::MRM4m: case X86II::MRM5m:
1001 case X86II::MRM6m: case X86II::MRM7m:
1002 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1004 EmitByte(BaseOpcode, CurByte, OS);
1005 EmitMemModRMByte(MI, CurOp, (TSFlags & X86II::FormMask)-X86II::MRM0m,
1006 TSFlags, CurByte, OS, Fixups);
1007 CurOp += X86::AddrNumOperands;
1010 EmitByte(BaseOpcode, CurByte, OS);
1011 EmitByte(0xC1, CurByte, OS);
1014 EmitByte(BaseOpcode, CurByte, OS);
1015 EmitByte(0xC2, CurByte, OS);
1018 EmitByte(BaseOpcode, CurByte, OS);
1019 EmitByte(0xC3, CurByte, OS);
1022 EmitByte(BaseOpcode, CurByte, OS);
1023 EmitByte(0xC4, CurByte, OS);
1026 EmitByte(BaseOpcode, CurByte, OS);
1027 EmitByte(0xC8, CurByte, OS);
1030 EmitByte(BaseOpcode, CurByte, OS);
1031 EmitByte(0xC9, CurByte, OS);
1034 EmitByte(BaseOpcode, CurByte, OS);
1035 EmitByte(0xE8, CurByte, OS);
1038 EmitByte(BaseOpcode, CurByte, OS);
1039 EmitByte(0xF0, CurByte, OS);
1042 EmitByte(BaseOpcode, CurByte, OS);
1043 EmitByte(0xF8, CurByte, OS);
1046 EmitByte(BaseOpcode, CurByte, OS);
1047 EmitByte(0xF9, CurByte, OS);
1050 EmitByte(BaseOpcode, CurByte, OS);
1051 EmitByte(0xD0, CurByte, OS);
1054 EmitByte(BaseOpcode, CurByte, OS);
1055 EmitByte(0xD1, CurByte, OS);
1059 // If there is a remaining operand, it must be a trailing immediate. Emit it
1060 // according to the right size for the instruction.
1061 if (CurOp != NumOps) {
1062 // The last source register of a 4 operand instruction in AVX is encoded
1063 // in bits[7:4] of a immediate byte, and bits[3:0] are ignored.
1064 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) {
1065 const MCOperand &MO = MI.getOperand(CurOp++);
1066 bool IsExtReg = X86II::isX86_64ExtendedReg(MO.getReg());
1067 unsigned RegNum = (IsExtReg ? (1 << 7) : 0);
1068 RegNum |= GetX86RegNum(MO) << 4;
1069 EmitImmediate(MCOperand::CreateImm(RegNum), 1, FK_Data_1, CurByte, OS,
1073 // FIXME: Is there a better way to know that we need a signed relocation?
1074 if (MI.getOpcode() == X86::ADD64ri32 ||
1075 MI.getOpcode() == X86::MOV64ri32 ||
1076 MI.getOpcode() == X86::MOV64mi32 ||
1077 MI.getOpcode() == X86::PUSH64i32)
1078 FixupKind = X86::reloc_signed_4byte;
1080 FixupKind = getImmFixupKind(TSFlags);
1081 EmitImmediate(MI.getOperand(CurOp++),
1082 X86II::getSizeOfImm(TSFlags), MCFixupKind(FixupKind),
1083 CurByte, OS, Fixups);
1087 if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
1088 EmitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS);
1092 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
1093 errs() << "Cannot encode all operands of: ";