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 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
390 bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
392 // VEX_R: opcode externsion equivalent to REX.R in
393 // 1's complement (inverted) form
395 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
396 // 0: Same as REX_R=1 (64 bit mode only)
398 unsigned char VEX_R = 0x1;
400 // VEX_X: equivalent to REX.X, only used when a
401 // register is used for index in SIB Byte.
403 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
404 // 0: Same as REX.X=1 (64-bit mode only)
405 unsigned char VEX_X = 0x1;
409 // 1: Same as REX_B=0 (ignored in 32-bit mode)
410 // 0: Same as REX_B=1 (64 bit mode only)
412 unsigned char VEX_B = 0x1;
414 // VEX_W: opcode specific (use like REX.W, or used for
415 // opcode extension, or ignored, depending on the opcode byte)
416 unsigned char VEX_W = 0;
418 // XOP_W: opcode specific, same bit as VEX_W, but used to
419 // swap operand 3 and 4 for FMA4 and XOP instructions
420 unsigned char XOP_W = 0;
422 // VEX_5M (VEX m-mmmmm field):
424 // 0b00000: Reserved for future use
425 // 0b00001: implied 0F leading opcode
426 // 0b00010: implied 0F 38 leading opcode bytes
427 // 0b00011: implied 0F 3A leading opcode bytes
428 // 0b00100-0b11111: Reserved for future use
430 unsigned char VEX_5M = 0x1;
432 // VEX_4V (VEX vvvv field): a register specifier
433 // (in 1's complement form) or 1111 if unused.
434 unsigned char VEX_4V = 0xf;
436 // VEX_L (Vector Length):
438 // 0: scalar or 128-bit vector
441 unsigned char VEX_L = 0;
443 // VEX_PP: opcode extension providing equivalent
444 // functionality of a SIMD prefix
451 unsigned char VEX_PP = 0;
453 // Encode the operand size opcode prefix as needed.
454 if (TSFlags & X86II::OpSize)
457 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_W)
460 if ((TSFlags >> X86II::VEXShift) & X86II::XOP_W)
463 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_L)
466 switch (TSFlags & X86II::Op0Mask) {
467 default: assert(0 && "Invalid prefix!");
468 case X86II::T8: // 0F 38
471 case X86II::TA: // 0F 3A
474 case X86II::T8XS: // F3 0F 38
478 case X86II::T8XD: // F2 0F 38
482 case X86II::TAXD: // F2 0F 3A
486 case X86II::XS: // F3 0F
489 case X86II::XD: // F2 0F
492 case X86II::A6: // Bypass: Not used by VEX
493 case X86II::A7: // Bypass: Not used by VEX
494 case X86II::TB: // Bypass: Not used by VEX
499 // Set the vector length to 256-bit if YMM0-YMM15 is used
500 for (unsigned i = 0; i != MI.getNumOperands(); ++i) {
501 if (!MI.getOperand(i).isReg())
503 unsigned SrcReg = MI.getOperand(i).getReg();
504 if (SrcReg >= X86::YMM0 && SrcReg <= X86::YMM15)
508 // Classify VEX_B, VEX_4V, VEX_R, VEX_X
510 switch (TSFlags & X86II::FormMask) {
511 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
512 case X86II::MRMDestMem: {
513 // MRMDestMem instructions forms:
514 // MemAddr, src1(ModR/M)
515 // MemAddr, src1(VEX_4V), src2(ModR/M)
516 // MemAddr, src1(ModR/M), imm8
518 if (X86II::isX86_64ExtendedReg(MI.getOperand(X86::AddrBaseReg).getReg()))
520 if (X86II::isX86_64ExtendedReg(MI.getOperand(X86::AddrIndexReg).getReg()))
523 CurOp = X86::AddrNumOperands;
525 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
527 const MCOperand &MO = MI.getOperand(CurOp);
528 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
532 case X86II::MRMSrcMem:
533 // MRMSrcMem instructions forms:
534 // src1(ModR/M), MemAddr
535 // src1(ModR/M), src2(VEX_4V), MemAddr
536 // src1(ModR/M), MemAddr, imm8
537 // src1(ModR/M), MemAddr, src2(VEX_I8IMM)
540 // dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
541 // dst(ModR/M.reg), src1(VEX_4V), src2(VEX_I8IMM), src3(ModR/M),
542 if (X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
546 VEX_4V = getVEXRegisterEncoding(MI, 1);
548 if (X86II::isX86_64ExtendedReg(
549 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
551 if (X86II::isX86_64ExtendedReg(
552 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
556 VEX_4V = getVEXRegisterEncoding(MI, X86::AddrNumOperands+1);
558 case X86II::MRM0m: case X86II::MRM1m:
559 case X86II::MRM2m: case X86II::MRM3m:
560 case X86II::MRM4m: case X86II::MRM5m:
561 case X86II::MRM6m: case X86II::MRM7m: {
562 // MRM[0-9]m instructions forms:
564 // src1(VEX_4V), MemAddr
566 VEX_4V = getVEXRegisterEncoding(MI, 0);
568 if (X86II::isX86_64ExtendedReg(
569 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
571 if (X86II::isX86_64ExtendedReg(
572 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
576 case X86II::MRMSrcReg:
577 // MRMSrcReg instructions forms:
578 // dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
579 // dst(ModR/M), src1(ModR/M)
580 // dst(ModR/M), src1(ModR/M), imm8
582 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
587 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
588 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
592 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
594 case X86II::MRMDestReg:
595 // MRMDestReg instructions forms:
596 // dst(ModR/M), src(ModR/M)
597 // dst(ModR/M), src(ModR/M), imm8
598 if (X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
600 if (X86II::isX86_64ExtendedReg(MI.getOperand(1).getReg()))
603 case X86II::MRM0r: case X86II::MRM1r:
604 case X86II::MRM2r: case X86II::MRM3r:
605 case X86II::MRM4r: case X86II::MRM5r:
606 case X86II::MRM6r: case X86II::MRM7r:
607 // MRM0r-MRM7r instructions forms:
608 // dst(VEX_4V), src(ModR/M), imm8
609 VEX_4V = getVEXRegisterEncoding(MI, 0);
610 if (X86II::isX86_64ExtendedReg(MI.getOperand(1).getReg()))
617 // Emit segment override opcode prefix as needed.
618 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
620 // VEX opcode prefix can have 2 or 3 bytes
623 // +-----+ +--------------+ +-------------------+
624 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
625 // +-----+ +--------------+ +-------------------+
627 // +-----+ +-------------------+
628 // | C5h | | R | vvvv | L | pp |
629 // +-----+ +-------------------+
631 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
633 if (VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) { // 2 byte VEX prefix
634 EmitByte(0xC5, CurByte, OS);
635 EmitByte(LastByte | (VEX_R << 7), CurByte, OS);
640 EmitByte(0xC4, CurByte, OS);
641 EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS);
642 EmitByte(LastByte | ((VEX_W | XOP_W) << 7), CurByte, OS);
645 /// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64
646 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
647 /// size, and 3) use of X86-64 extended registers.
648 static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags,
649 const MCInstrDesc &Desc) {
651 if (TSFlags & X86II::REX_W)
652 REX |= 1 << 3; // set REX.W
654 if (MI.getNumOperands() == 0) return REX;
656 unsigned NumOps = MI.getNumOperands();
657 // FIXME: MCInst should explicitize the two-addrness.
658 bool isTwoAddr = NumOps > 1 &&
659 Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1;
661 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
662 unsigned i = isTwoAddr ? 1 : 0;
663 for (; i != NumOps; ++i) {
664 const MCOperand &MO = MI.getOperand(i);
665 if (!MO.isReg()) continue;
666 unsigned Reg = MO.getReg();
667 if (!X86II::isX86_64NonExtLowByteReg(Reg)) continue;
668 // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything
669 // that returns non-zero.
670 REX |= 0x40; // REX fixed encoding prefix
674 switch (TSFlags & X86II::FormMask) {
675 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!");
676 case X86II::MRMSrcReg:
677 if (MI.getOperand(0).isReg() &&
678 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
679 REX |= 1 << 2; // set REX.R
680 i = isTwoAddr ? 2 : 1;
681 for (; i != NumOps; ++i) {
682 const MCOperand &MO = MI.getOperand(i);
683 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
684 REX |= 1 << 0; // set REX.B
687 case X86II::MRMSrcMem: {
688 if (MI.getOperand(0).isReg() &&
689 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
690 REX |= 1 << 2; // set REX.R
692 i = isTwoAddr ? 2 : 1;
693 for (; i != NumOps; ++i) {
694 const MCOperand &MO = MI.getOperand(i);
696 if (X86II::isX86_64ExtendedReg(MO.getReg()))
697 REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1)
703 case X86II::MRM0m: case X86II::MRM1m:
704 case X86II::MRM2m: case X86II::MRM3m:
705 case X86II::MRM4m: case X86II::MRM5m:
706 case X86II::MRM6m: case X86II::MRM7m:
707 case X86II::MRMDestMem: {
708 unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands);
709 i = isTwoAddr ? 1 : 0;
710 if (NumOps > e && MI.getOperand(e).isReg() &&
711 X86II::isX86_64ExtendedReg(MI.getOperand(e).getReg()))
712 REX |= 1 << 2; // set REX.R
714 for (; i != e; ++i) {
715 const MCOperand &MO = MI.getOperand(i);
717 if (X86II::isX86_64ExtendedReg(MO.getReg()))
718 REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1)
725 if (MI.getOperand(0).isReg() &&
726 X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg()))
727 REX |= 1 << 0; // set REX.B
728 i = isTwoAddr ? 2 : 1;
729 for (unsigned e = NumOps; i != e; ++i) {
730 const MCOperand &MO = MI.getOperand(i);
731 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
732 REX |= 1 << 2; // set REX.R
739 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
740 void X86MCCodeEmitter::EmitSegmentOverridePrefix(uint64_t TSFlags,
741 unsigned &CurByte, int MemOperand,
743 raw_ostream &OS) const {
744 switch (TSFlags & X86II::SegOvrMask) {
745 default: assert(0 && "Invalid segment!");
747 // No segment override, check for explicit one on memory operand.
748 if (MemOperand != -1) { // If the instruction has a memory operand.
749 switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
750 default: assert(0 && "Unknown segment register!");
752 case X86::CS: EmitByte(0x2E, CurByte, OS); break;
753 case X86::SS: EmitByte(0x36, CurByte, OS); break;
754 case X86::DS: EmitByte(0x3E, CurByte, OS); break;
755 case X86::ES: EmitByte(0x26, CurByte, OS); break;
756 case X86::FS: EmitByte(0x64, CurByte, OS); break;
757 case X86::GS: EmitByte(0x65, CurByte, OS); break;
762 EmitByte(0x64, CurByte, OS);
765 EmitByte(0x65, CurByte, OS);
770 /// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode.
772 /// MemOperand is the operand # of the start of a memory operand if present. If
773 /// Not present, it is -1.
774 void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte,
775 int MemOperand, const MCInst &MI,
776 const MCInstrDesc &Desc,
777 raw_ostream &OS) const {
779 // Emit the lock opcode prefix as needed.
780 if (TSFlags & X86II::LOCK)
781 EmitByte(0xF0, CurByte, OS);
783 // Emit segment override opcode prefix as needed.
784 EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS);
786 // Emit the repeat opcode prefix as needed.
787 if ((TSFlags & X86II::Op0Mask) == X86II::REP)
788 EmitByte(0xF3, CurByte, OS);
790 // Emit the address size opcode prefix as needed.
791 if ((TSFlags & X86II::AdSize) ||
792 (MemOperand != -1 && is64BitMode() && Is32BitMemOperand(MI, MemOperand)))
793 EmitByte(0x67, CurByte, OS);
795 // Emit the operand size opcode prefix as needed.
796 if (TSFlags & X86II::OpSize)
797 EmitByte(0x66, CurByte, OS);
799 bool Need0FPrefix = false;
800 switch (TSFlags & X86II::Op0Mask) {
801 default: assert(0 && "Invalid prefix!");
802 case 0: break; // No prefix!
803 case X86II::REP: break; // already handled.
804 case X86II::TB: // Two-byte opcode prefix
805 case X86II::T8: // 0F 38
806 case X86II::TA: // 0F 3A
807 case X86II::A6: // 0F A6
808 case X86II::A7: // 0F A7
811 case X86II::T8XS: // F3 0F 38
812 EmitByte(0xF3, CurByte, OS);
815 case X86II::T8XD: // F2 0F 38
816 EmitByte(0xF2, CurByte, OS);
819 case X86II::TAXD: // F2 0F 3A
820 EmitByte(0xF2, CurByte, OS);
823 case X86II::XS: // F3 0F
824 EmitByte(0xF3, CurByte, OS);
827 case X86II::XD: // F2 0F
828 EmitByte(0xF2, CurByte, OS);
831 case X86II::D8: EmitByte(0xD8, CurByte, OS); break;
832 case X86II::D9: EmitByte(0xD9, CurByte, OS); break;
833 case X86II::DA: EmitByte(0xDA, CurByte, OS); break;
834 case X86II::DB: EmitByte(0xDB, CurByte, OS); break;
835 case X86II::DC: EmitByte(0xDC, CurByte, OS); break;
836 case X86II::DD: EmitByte(0xDD, CurByte, OS); break;
837 case X86II::DE: EmitByte(0xDE, CurByte, OS); break;
838 case X86II::DF: EmitByte(0xDF, CurByte, OS); break;
841 // Handle REX prefix.
842 // FIXME: Can this come before F2 etc to simplify emission?
844 if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc))
845 EmitByte(0x40 | REX, CurByte, OS);
848 // 0x0F escape code must be emitted just before the opcode.
850 EmitByte(0x0F, CurByte, OS);
852 // FIXME: Pull this up into previous switch if REX can be moved earlier.
853 switch (TSFlags & X86II::Op0Mask) {
854 case X86II::T8XS: // F3 0F 38
855 case X86II::T8XD: // F2 0F 38
856 case X86II::T8: // 0F 38
857 EmitByte(0x38, CurByte, OS);
859 case X86II::TAXD: // F2 0F 3A
860 case X86II::TA: // 0F 3A
861 EmitByte(0x3A, CurByte, OS);
863 case X86II::A6: // 0F A6
864 EmitByte(0xA6, CurByte, OS);
866 case X86II::A7: // 0F A7
867 EmitByte(0xA7, CurByte, OS);
872 void X86MCCodeEmitter::
873 EncodeInstruction(const MCInst &MI, raw_ostream &OS,
874 SmallVectorImpl<MCFixup> &Fixups) const {
875 unsigned Opcode = MI.getOpcode();
876 const MCInstrDesc &Desc = MCII.get(Opcode);
877 uint64_t TSFlags = Desc.TSFlags;
879 // Pseudo instructions don't get encoded.
880 if ((TSFlags & X86II::FormMask) == X86II::Pseudo)
883 // If this is a two-address instruction, skip one of the register operands.
884 // FIXME: This should be handled during MCInst lowering.
885 unsigned NumOps = Desc.getNumOperands();
887 if (NumOps > 1 && Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1)
889 else if (NumOps > 2 && Desc.getOperandConstraint(NumOps-1, MCOI::TIED_TO)== 0)
890 // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32
893 // Keep track of the current byte being emitted.
894 unsigned CurByte = 0;
896 // Is this instruction encoded using the AVX VEX prefix?
897 bool HasVEXPrefix = (TSFlags >> X86II::VEXShift) & X86II::VEX;
899 // It uses the VEX.VVVV field?
900 bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
901 bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
902 bool HasXOP_W = (TSFlags >> X86II::VEXShift) & X86II::XOP_W;
903 unsigned XOP_W_I8IMMOperand = 2;
905 // Determine where the memory operand starts, if present.
906 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags, Opcode);
907 if (MemoryOperand != -1) MemoryOperand += CurOp;
910 EmitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
912 EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS);
914 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags);
916 if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
917 BaseOpcode = 0x0F; // Weird 3DNow! encoding.
919 unsigned SrcRegNum = 0;
920 switch (TSFlags & X86II::FormMask) {
921 case X86II::MRMInitReg:
922 assert(0 && "FIXME: Remove this form when the JIT moves to MCCodeEmitter!");
923 default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n";
924 assert(0 && "Unknown FormMask value in X86MCCodeEmitter!");
926 assert(0 && "Pseudo instruction shouldn't be emitted");
928 EmitByte(BaseOpcode, CurByte, OS);
930 case X86II::RawFrmImm8:
931 EmitByte(BaseOpcode, CurByte, OS);
932 EmitImmediate(MI.getOperand(CurOp++),
933 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
934 CurByte, OS, Fixups);
935 EmitImmediate(MI.getOperand(CurOp++), 1, FK_Data_1, CurByte, OS, Fixups);
937 case X86II::RawFrmImm16:
938 EmitByte(BaseOpcode, CurByte, OS);
939 EmitImmediate(MI.getOperand(CurOp++),
940 X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags),
941 CurByte, OS, Fixups);
942 EmitImmediate(MI.getOperand(CurOp++), 2, FK_Data_2, CurByte, OS, Fixups);
945 case X86II::AddRegFrm:
946 EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS);
949 case X86II::MRMDestReg:
950 EmitByte(BaseOpcode, CurByte, OS);
951 EmitRegModRMByte(MI.getOperand(CurOp),
952 GetX86RegNum(MI.getOperand(CurOp+1)), CurByte, OS);
956 case X86II::MRMDestMem:
957 EmitByte(BaseOpcode, CurByte, OS);
958 SrcRegNum = CurOp + X86::AddrNumOperands;
960 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
963 EmitMemModRMByte(MI, CurOp,
964 GetX86RegNum(MI.getOperand(SrcRegNum)),
965 TSFlags, CurByte, OS, Fixups);
966 CurOp = SrcRegNum + 1;
969 case X86II::MRMSrcReg:
970 EmitByte(BaseOpcode, CurByte, OS);
971 SrcRegNum = CurOp + 1;
973 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
976 // GAS sets the XOP_W even with register operands, we want to match this.
977 // XOP_W is ignored, so there is no swapping of the operands
978 XOP_W_I8IMMOperand = 3;
980 EmitRegModRMByte(MI.getOperand(SrcRegNum),
981 GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS);
982 CurOp = SrcRegNum + 1;
987 case X86II::MRMSrcMem: {
988 int AddrOperands = X86::AddrNumOperands;
989 unsigned FirstMemOp = CurOp+1;
992 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
994 if(HasXOP_W) // Skip second register source (encoded in I8IMM)
997 EmitByte(BaseOpcode, CurByte, OS);
999 EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)),
1000 TSFlags, CurByte, OS, Fixups);
1002 CurOp = NumOps - 1; // We have consumed all except one operand (third)
1004 CurOp += AddrOperands + 1;
1011 case X86II::MRM0r: case X86II::MRM1r:
1012 case X86II::MRM2r: case X86II::MRM3r:
1013 case X86II::MRM4r: case X86II::MRM5r:
1014 case X86II::MRM6r: case X86II::MRM7r:
1015 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1017 EmitByte(BaseOpcode, CurByte, OS);
1018 EmitRegModRMByte(MI.getOperand(CurOp++),
1019 (TSFlags & X86II::FormMask)-X86II::MRM0r,
1022 case X86II::MRM0m: case X86II::MRM1m:
1023 case X86II::MRM2m: case X86II::MRM3m:
1024 case X86II::MRM4m: case X86II::MRM5m:
1025 case X86II::MRM6m: case X86II::MRM7m:
1026 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1028 EmitByte(BaseOpcode, CurByte, OS);
1029 EmitMemModRMByte(MI, CurOp, (TSFlags & X86II::FormMask)-X86II::MRM0m,
1030 TSFlags, CurByte, OS, Fixups);
1031 CurOp += X86::AddrNumOperands;
1034 EmitByte(BaseOpcode, CurByte, OS);
1035 EmitByte(0xC1, CurByte, OS);
1038 EmitByte(BaseOpcode, CurByte, OS);
1039 EmitByte(0xC2, CurByte, OS);
1042 EmitByte(BaseOpcode, CurByte, OS);
1043 EmitByte(0xC3, CurByte, OS);
1046 EmitByte(BaseOpcode, CurByte, OS);
1047 EmitByte(0xC4, CurByte, OS);
1050 EmitByte(BaseOpcode, CurByte, OS);
1051 EmitByte(0xC8, CurByte, OS);
1054 EmitByte(BaseOpcode, CurByte, OS);
1055 EmitByte(0xC9, CurByte, OS);
1058 EmitByte(BaseOpcode, CurByte, OS);
1059 EmitByte(0xE8, CurByte, OS);
1062 EmitByte(BaseOpcode, CurByte, OS);
1063 EmitByte(0xF0, CurByte, OS);
1066 EmitByte(BaseOpcode, CurByte, OS);
1067 EmitByte(0xF8, CurByte, OS);
1070 EmitByte(BaseOpcode, CurByte, OS);
1071 EmitByte(0xF9, CurByte, OS);
1074 EmitByte(BaseOpcode, CurByte, OS);
1075 EmitByte(0xD0, CurByte, OS);
1078 EmitByte(BaseOpcode, CurByte, OS);
1079 EmitByte(0xD1, CurByte, OS);
1083 // If there is a remaining operand, it must be a trailing immediate. Emit it
1084 // according to the right size for the instruction.
1085 if (CurOp != NumOps) {
1086 // The last source register of a 4 operand instruction in AVX is encoded
1087 // in bits[7:4] of a immediate byte, and bits[3:0] are ignored.
1088 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) {
1089 const MCOperand &MO = MI.getOperand(HasXOP_W ? XOP_W_I8IMMOperand
1092 bool IsExtReg = X86II::isX86_64ExtendedReg(MO.getReg());
1093 unsigned RegNum = (IsExtReg ? (1 << 7) : 0);
1094 RegNum |= GetX86RegNum(MO) << 4;
1095 EmitImmediate(MCOperand::CreateImm(RegNum), 1, FK_Data_1, CurByte, OS,
1099 // FIXME: Is there a better way to know that we need a signed relocation?
1100 if (MI.getOpcode() == X86::ADD64ri32 ||
1101 MI.getOpcode() == X86::MOV64ri32 ||
1102 MI.getOpcode() == X86::MOV64mi32 ||
1103 MI.getOpcode() == X86::PUSH64i32)
1104 FixupKind = X86::reloc_signed_4byte;
1106 FixupKind = getImmFixupKind(TSFlags);
1107 EmitImmediate(MI.getOperand(CurOp++),
1108 X86II::getSizeOfImm(TSFlags), MCFixupKind(FixupKind),
1109 CurByte, OS, Fixups);
1113 if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode)
1114 EmitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS);
1118 if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) {
1119 errs() << "Cannot encode all operands of: ";