1 //===-- X86CodeEmitter.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 contains the pass that transforms the X86 machine instructions into
11 // relocatable machine code.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-emitter"
17 #include "X86InstrInfo.h"
18 #include "X86JITInfo.h"
19 #include "X86Relocations.h"
20 #include "X86Subtarget.h"
21 #include "X86TargetMachine.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/JITCodeEmitter.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineInstr.h"
26 #include "llvm/CodeGen/MachineModuleInfo.h"
27 #include "llvm/CodeGen/Passes.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/MC/MCCodeEmitter.h"
30 #include "llvm/MC/MCExpr.h"
31 #include "llvm/MC/MCInst.h"
32 #include "llvm/PassManager.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 STATISTIC(NumEmitted, "Number of machine instructions emitted");
42 template<class CodeEmitter>
43 class Emitter : public MachineFunctionPass {
44 const X86InstrInfo *II;
48 MachineModuleInfo *MMI;
49 intptr_t PICBaseOffset;
54 explicit Emitter(X86TargetMachine &tm, CodeEmitter &mce)
55 : MachineFunctionPass(ID), II(0), TD(0), TM(tm),
56 MCE(mce), PICBaseOffset(0), Is64BitMode(false),
57 IsPIC(TM.getRelocationModel() == Reloc::PIC_) {}
59 bool runOnMachineFunction(MachineFunction &MF);
61 virtual const char *getPassName() const {
62 return "X86 Machine Code Emitter";
65 void emitOpcodePrefix(uint64_t TSFlags, int MemOperand,
66 const MachineInstr &MI,
67 const MCInstrDesc *Desc) const;
69 void emitVEXOpcodePrefix(uint64_t TSFlags, int MemOperand,
70 const MachineInstr &MI,
71 const MCInstrDesc *Desc) const;
73 void emitSegmentOverridePrefix(uint64_t TSFlags,
75 const MachineInstr &MI) const;
77 void emitInstruction(MachineInstr &MI, const MCInstrDesc *Desc);
79 void getAnalysisUsage(AnalysisUsage &AU) const {
81 AU.addRequired<MachineModuleInfo>();
82 MachineFunctionPass::getAnalysisUsage(AU);
86 void emitPCRelativeBlockAddress(MachineBasicBlock *MBB);
87 void emitGlobalAddress(const GlobalValue *GV, unsigned Reloc,
88 intptr_t Disp = 0, intptr_t PCAdj = 0,
89 bool Indirect = false);
90 void emitExternalSymbolAddress(const char *ES, unsigned Reloc);
91 void emitConstPoolAddress(unsigned CPI, unsigned Reloc, intptr_t Disp = 0,
93 void emitJumpTableAddress(unsigned JTI, unsigned Reloc,
96 void emitDisplacementField(const MachineOperand *RelocOp, int DispVal,
97 intptr_t Adj = 0, bool IsPCRel = true);
99 void emitRegModRMByte(unsigned ModRMReg, unsigned RegOpcodeField);
100 void emitRegModRMByte(unsigned RegOpcodeField);
101 void emitSIBByte(unsigned SS, unsigned Index, unsigned Base);
102 void emitConstant(uint64_t Val, unsigned Size);
104 void emitMemModRMByte(const MachineInstr &MI,
105 unsigned Op, unsigned RegOpcodeField,
108 unsigned getX86RegNum(unsigned RegNo) const {
109 const TargetRegisterInfo *TRI = TM.getRegisterInfo();
110 return TRI->getEncodingValue(RegNo) & 0x7;
113 unsigned char getVEXRegisterEncoding(const MachineInstr &MI,
114 unsigned OpNum) const;
117 template<class CodeEmitter>
118 char Emitter<CodeEmitter>::ID = 0;
119 } // end anonymous namespace.
121 /// createX86CodeEmitterPass - Return a pass that emits the collected X86 code
122 /// to the specified JITCodeEmitter object.
123 FunctionPass *llvm::createX86JITCodeEmitterPass(X86TargetMachine &TM,
124 JITCodeEmitter &JCE) {
125 return new Emitter<JITCodeEmitter>(TM, JCE);
128 template<class CodeEmitter>
129 bool Emitter<CodeEmitter>::runOnMachineFunction(MachineFunction &MF) {
130 MMI = &getAnalysis<MachineModuleInfo>();
131 MCE.setModuleInfo(MMI);
133 II = TM.getInstrInfo();
134 TD = TM.getDataLayout();
135 Is64BitMode = TM.getSubtarget<X86Subtarget>().is64Bit();
136 IsPIC = TM.getRelocationModel() == Reloc::PIC_;
139 DEBUG(dbgs() << "JITTing function '" << MF.getName() << "'\n");
140 MCE.startFunction(MF);
141 for (MachineFunction::iterator MBB = MF.begin(), E = MF.end();
143 MCE.StartMachineBasicBlock(MBB);
144 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
146 const MCInstrDesc &Desc = I->getDesc();
147 emitInstruction(*I, &Desc);
148 // MOVPC32r is basically a call plus a pop instruction.
149 if (Desc.getOpcode() == X86::MOVPC32r)
150 emitInstruction(*I, &II->get(X86::POP32r));
151 ++NumEmitted; // Keep track of the # of mi's emitted
154 } while (MCE.finishFunction(MF));
159 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
160 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
161 /// size, and 3) use of X86-64 extended registers.
162 static unsigned determineREX(const MachineInstr &MI) {
164 const MCInstrDesc &Desc = MI.getDesc();
166 // Pseudo instructions do not need REX prefix byte.
167 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
169 if (Desc.TSFlags & X86II::REX_W)
172 unsigned NumOps = Desc.getNumOperands();
174 bool isTwoAddr = NumOps > 1 &&
175 Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1;
177 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
178 unsigned i = isTwoAddr ? 1 : 0;
179 for (unsigned e = NumOps; i != e; ++i) {
180 const MachineOperand& MO = MI.getOperand(i);
182 unsigned Reg = MO.getReg();
183 if (X86II::isX86_64NonExtLowByteReg(Reg))
188 switch (Desc.TSFlags & X86II::FormMask) {
189 case X86II::MRMSrcReg: {
190 if (X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0)))
192 i = isTwoAddr ? 2 : 1;
193 for (unsigned e = NumOps; i != e; ++i) {
194 const MachineOperand& MO = MI.getOperand(i);
195 if (X86InstrInfo::isX86_64ExtendedReg(MO))
200 case X86II::MRMSrcMem: {
201 if (X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0)))
204 i = isTwoAddr ? 2 : 1;
205 for (; i != NumOps; ++i) {
206 const MachineOperand& MO = MI.getOperand(i);
208 if (X86InstrInfo::isX86_64ExtendedReg(MO))
215 case X86II::MRM0m: case X86II::MRM1m:
216 case X86II::MRM2m: case X86II::MRM3m:
217 case X86II::MRM4m: case X86II::MRM5m:
218 case X86II::MRM6m: case X86II::MRM7m:
219 case X86II::MRMDestMem: {
220 unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands);
221 i = isTwoAddr ? 1 : 0;
222 if (NumOps > e && X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(e)))
225 for (; i != e; ++i) {
226 const MachineOperand& MO = MI.getOperand(i);
228 if (X86InstrInfo::isX86_64ExtendedReg(MO))
236 if (X86InstrInfo::isX86_64ExtendedReg(MI.getOperand(0)))
238 i = isTwoAddr ? 2 : 1;
239 for (unsigned e = NumOps; i != e; ++i) {
240 const MachineOperand& MO = MI.getOperand(i);
241 if (X86InstrInfo::isX86_64ExtendedReg(MO))
252 /// emitPCRelativeBlockAddress - This method keeps track of the information
253 /// necessary to resolve the address of this block later and emits a dummy
256 template<class CodeEmitter>
257 void Emitter<CodeEmitter>::emitPCRelativeBlockAddress(MachineBasicBlock *MBB) {
258 // Remember where this reference was and where it is to so we can
259 // deal with it later.
260 MCE.addRelocation(MachineRelocation::getBB(MCE.getCurrentPCOffset(),
261 X86::reloc_pcrel_word, MBB));
265 /// emitGlobalAddress - Emit the specified address to the code stream assuming
266 /// this is part of a "take the address of a global" instruction.
268 template<class CodeEmitter>
269 void Emitter<CodeEmitter>::emitGlobalAddress(const GlobalValue *GV,
271 intptr_t Disp /* = 0 */,
272 intptr_t PCAdj /* = 0 */,
273 bool Indirect /* = false */) {
274 intptr_t RelocCST = Disp;
275 if (Reloc == X86::reloc_picrel_word)
276 RelocCST = PICBaseOffset;
277 else if (Reloc == X86::reloc_pcrel_word)
279 MachineRelocation MR = Indirect
280 ? MachineRelocation::getIndirectSymbol(MCE.getCurrentPCOffset(), Reloc,
281 const_cast<GlobalValue *>(GV),
283 : MachineRelocation::getGV(MCE.getCurrentPCOffset(), Reloc,
284 const_cast<GlobalValue *>(GV), RelocCST, false);
285 MCE.addRelocation(MR);
286 // The relocated value will be added to the displacement
287 if (Reloc == X86::reloc_absolute_dword)
288 MCE.emitDWordLE(Disp);
290 MCE.emitWordLE((int32_t)Disp);
293 /// emitExternalSymbolAddress - Arrange for the address of an external symbol to
294 /// be emitted to the current location in the function, and allow it to be PC
296 template<class CodeEmitter>
297 void Emitter<CodeEmitter>::emitExternalSymbolAddress(const char *ES,
299 intptr_t RelocCST = (Reloc == X86::reloc_picrel_word) ? PICBaseOffset : 0;
301 // X86 never needs stubs because instruction selection will always pick
302 // an instruction sequence that is large enough to hold any address
304 // (see X86ISelLowering.cpp, near 2039: X86TargetLowering::LowerCall)
305 bool NeedStub = false;
306 MCE.addRelocation(MachineRelocation::getExtSym(MCE.getCurrentPCOffset(),
309 if (Reloc == X86::reloc_absolute_dword)
315 /// emitConstPoolAddress - Arrange for the address of an constant pool
316 /// to be emitted to the current location in the function, and allow it to be PC
318 template<class CodeEmitter>
319 void Emitter<CodeEmitter>::emitConstPoolAddress(unsigned CPI, unsigned Reloc,
320 intptr_t Disp /* = 0 */,
321 intptr_t PCAdj /* = 0 */) {
322 intptr_t RelocCST = 0;
323 if (Reloc == X86::reloc_picrel_word)
324 RelocCST = PICBaseOffset;
325 else if (Reloc == X86::reloc_pcrel_word)
327 MCE.addRelocation(MachineRelocation::getConstPool(MCE.getCurrentPCOffset(),
328 Reloc, CPI, RelocCST));
329 // The relocated value will be added to the displacement
330 if (Reloc == X86::reloc_absolute_dword)
331 MCE.emitDWordLE(Disp);
333 MCE.emitWordLE((int32_t)Disp);
336 /// emitJumpTableAddress - Arrange for the address of a jump table to
337 /// be emitted to the current location in the function, and allow it to be PC
339 template<class CodeEmitter>
340 void Emitter<CodeEmitter>::emitJumpTableAddress(unsigned JTI, unsigned Reloc,
341 intptr_t PCAdj /* = 0 */) {
342 intptr_t RelocCST = 0;
343 if (Reloc == X86::reloc_picrel_word)
344 RelocCST = PICBaseOffset;
345 else if (Reloc == X86::reloc_pcrel_word)
347 MCE.addRelocation(MachineRelocation::getJumpTable(MCE.getCurrentPCOffset(),
348 Reloc, JTI, RelocCST));
349 // The relocated value will be added to the displacement
350 if (Reloc == X86::reloc_absolute_dword)
356 inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode,
358 assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!");
359 return RM | (RegOpcode << 3) | (Mod << 6);
362 template<class CodeEmitter>
363 void Emitter<CodeEmitter>::emitRegModRMByte(unsigned ModRMReg,
364 unsigned RegOpcodeFld){
365 MCE.emitByte(ModRMByte(3, RegOpcodeFld, getX86RegNum(ModRMReg)));
368 template<class CodeEmitter>
369 void Emitter<CodeEmitter>::emitRegModRMByte(unsigned RegOpcodeFld) {
370 MCE.emitByte(ModRMByte(3, RegOpcodeFld, 0));
373 template<class CodeEmitter>
374 void Emitter<CodeEmitter>::emitSIBByte(unsigned SS,
377 // SIB byte is in the same format as the ModRMByte...
378 MCE.emitByte(ModRMByte(SS, Index, Base));
381 template<class CodeEmitter>
382 void Emitter<CodeEmitter>::emitConstant(uint64_t Val, unsigned Size) {
383 // Output the constant in little endian byte order...
384 for (unsigned i = 0; i != Size; ++i) {
385 MCE.emitByte(Val & 255);
390 /// isDisp8 - Return true if this signed displacement fits in a 8-bit
391 /// sign-extended field.
392 static bool isDisp8(int Value) {
393 return Value == (signed char)Value;
396 static bool gvNeedsNonLazyPtr(const MachineOperand &GVOp,
397 const TargetMachine &TM) {
398 // For Darwin-64, simulate the linktime GOT by using the same non-lazy-pointer
399 // mechanism as 32-bit mode.
400 if (TM.getSubtarget<X86Subtarget>().is64Bit() &&
401 !TM.getSubtarget<X86Subtarget>().isTargetDarwin())
404 // Return true if this is a reference to a stub containing the address of the
405 // global, not the global itself.
406 return isGlobalStubReference(GVOp.getTargetFlags());
409 template<class CodeEmitter>
410 void Emitter<CodeEmitter>::emitDisplacementField(const MachineOperand *RelocOp,
412 intptr_t Adj /* = 0 */,
413 bool IsPCRel /* = true */) {
414 // If this is a simple integer displacement that doesn't require a relocation,
417 emitConstant(DispVal, 4);
421 // Otherwise, this is something that requires a relocation. Emit it as such
423 unsigned RelocType = Is64BitMode ?
424 (IsPCRel ? X86::reloc_pcrel_word : X86::reloc_absolute_word_sext)
425 : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word);
426 if (RelocOp->isGlobal()) {
427 // In 64-bit static small code model, we could potentially emit absolute.
428 // But it's probably not beneficial. If the MCE supports using RIP directly
429 // do it, otherwise fallback to absolute (this is determined by IsPCRel).
430 // 89 05 00 00 00 00 mov %eax,0(%rip) # PC-relative
431 // 89 04 25 00 00 00 00 mov %eax,0x0 # Absolute
432 bool Indirect = gvNeedsNonLazyPtr(*RelocOp, TM);
433 emitGlobalAddress(RelocOp->getGlobal(), RelocType, RelocOp->getOffset(),
435 } else if (RelocOp->isSymbol()) {
436 emitExternalSymbolAddress(RelocOp->getSymbolName(), RelocType);
437 } else if (RelocOp->isCPI()) {
438 emitConstPoolAddress(RelocOp->getIndex(), RelocType,
439 RelocOp->getOffset(), Adj);
441 assert(RelocOp->isJTI() && "Unexpected machine operand!");
442 emitJumpTableAddress(RelocOp->getIndex(), RelocType, Adj);
446 template<class CodeEmitter>
447 void Emitter<CodeEmitter>::emitMemModRMByte(const MachineInstr &MI,
448 unsigned Op,unsigned RegOpcodeField,
450 const MachineOperand &Op3 = MI.getOperand(Op+3);
452 const MachineOperand *DispForReloc = 0;
454 // Figure out what sort of displacement we have to handle here.
455 if (Op3.isGlobal()) {
457 } else if (Op3.isSymbol()) {
459 } else if (Op3.isCPI()) {
460 if (!MCE.earlyResolveAddresses() || Is64BitMode || IsPIC) {
463 DispVal += MCE.getConstantPoolEntryAddress(Op3.getIndex());
464 DispVal += Op3.getOffset();
466 } else if (Op3.isJTI()) {
467 if (!MCE.earlyResolveAddresses() || Is64BitMode || IsPIC) {
470 DispVal += MCE.getJumpTableEntryAddress(Op3.getIndex());
473 DispVal = Op3.getImm();
476 const MachineOperand &Base = MI.getOperand(Op);
477 const MachineOperand &Scale = MI.getOperand(Op+1);
478 const MachineOperand &IndexReg = MI.getOperand(Op+2);
480 unsigned BaseReg = Base.getReg();
482 // Handle %rip relative addressing.
483 if (BaseReg == X86::RIP ||
484 (Is64BitMode && DispForReloc)) { // [disp32+RIP] in X86-64 mode
485 assert(IndexReg.getReg() == 0 && Is64BitMode &&
486 "Invalid rip-relative address");
487 MCE.emitByte(ModRMByte(0, RegOpcodeField, 5));
488 emitDisplacementField(DispForReloc, DispVal, PCAdj, true);
492 // Indicate that the displacement will use an pcrel or absolute reference
493 // by default. MCEs able to resolve addresses on-the-fly use pcrel by default
494 // while others, unless explicit asked to use RIP, use absolute references.
495 bool IsPCRel = MCE.earlyResolveAddresses() ? true : false;
497 // Is a SIB byte needed?
498 // If no BaseReg, issue a RIP relative instruction only if the MCE can
499 // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table
500 // 2-7) and absolute references.
501 unsigned BaseRegNo = -1U;
502 if (BaseReg != 0 && BaseReg != X86::RIP)
503 BaseRegNo = getX86RegNum(BaseReg);
505 if (// The SIB byte must be used if there is an index register.
506 IndexReg.getReg() == 0 &&
507 // The SIB byte must be used if the base is ESP/RSP/R12, all of which
508 // encode to an R/M value of 4, which indicates that a SIB byte is
510 BaseRegNo != N86::ESP &&
511 // If there is no base register and we're in 64-bit mode, we need a SIB
512 // byte to emit an addr that is just 'disp32' (the non-RIP relative form).
513 (!Is64BitMode || BaseReg != 0)) {
514 if (BaseReg == 0 || // [disp32] in X86-32 mode
515 BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode
516 MCE.emitByte(ModRMByte(0, RegOpcodeField, 5));
517 emitDisplacementField(DispForReloc, DispVal, PCAdj, true);
521 // If the base is not EBP/ESP and there is no displacement, use simple
522 // indirect register encoding, this handles addresses like [EAX]. The
523 // encoding for [EBP] with no displacement means [disp32] so we handle it
524 // by emitting a displacement of 0 below.
525 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
526 MCE.emitByte(ModRMByte(0, RegOpcodeField, BaseRegNo));
530 // Otherwise, if the displacement fits in a byte, encode as [REG+disp8].
531 if (!DispForReloc && isDisp8(DispVal)) {
532 MCE.emitByte(ModRMByte(1, RegOpcodeField, BaseRegNo));
533 emitConstant(DispVal, 1);
537 // Otherwise, emit the most general non-SIB encoding: [REG+disp32]
538 MCE.emitByte(ModRMByte(2, RegOpcodeField, BaseRegNo));
539 emitDisplacementField(DispForReloc, DispVal, PCAdj, IsPCRel);
543 // Otherwise we need a SIB byte, so start by outputting the ModR/M byte first.
544 assert(IndexReg.getReg() != X86::ESP &&
545 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
547 bool ForceDisp32 = false;
548 bool ForceDisp8 = false;
550 // If there is no base register, we emit the special case SIB byte with
551 // MOD=0, BASE=4, to JUST get the index, scale, and displacement.
552 MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
554 } else if (DispForReloc) {
555 // Emit the normal disp32 encoding.
556 MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
558 } else if (DispVal == 0 && BaseRegNo != N86::EBP) {
559 // Emit no displacement ModR/M byte
560 MCE.emitByte(ModRMByte(0, RegOpcodeField, 4));
561 } else if (isDisp8(DispVal)) {
562 // Emit the disp8 encoding...
563 MCE.emitByte(ModRMByte(1, RegOpcodeField, 4));
564 ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP
566 // Emit the normal disp32 encoding...
567 MCE.emitByte(ModRMByte(2, RegOpcodeField, 4));
570 // Calculate what the SS field value should be...
571 static const unsigned SSTable[] = { ~0U, 0, 1, ~0U, 2, ~0U, ~0U, ~0U, 3 };
572 unsigned SS = SSTable[Scale.getImm()];
575 // Handle the SIB byte for the case where there is no base, see Intel
576 // Manual 2A, table 2-7. The displacement has already been output.
578 if (IndexReg.getReg())
579 IndexRegNo = getX86RegNum(IndexReg.getReg());
580 else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5)
582 emitSIBByte(SS, IndexRegNo, 5);
584 unsigned BaseRegNo = getX86RegNum(BaseReg);
586 if (IndexReg.getReg())
587 IndexRegNo = getX86RegNum(IndexReg.getReg());
589 IndexRegNo = 4; // For example [ESP+1*<noreg>+4]
590 emitSIBByte(SS, IndexRegNo, BaseRegNo);
593 // Do we need to output a displacement?
595 emitConstant(DispVal, 1);
596 } else if (DispVal != 0 || ForceDisp32) {
597 emitDisplacementField(DispForReloc, DispVal, PCAdj, IsPCRel);
601 static const MCInstrDesc *UpdateOp(MachineInstr &MI, const X86InstrInfo *II,
603 const MCInstrDesc *Desc = &II->get(Opcode);
608 /// Is16BitMemOperand - Return true if the specified instruction has
609 /// a 16-bit memory operand. Op specifies the operand # of the memoperand.
610 static bool Is16BitMemOperand(const MachineInstr &MI, unsigned Op) {
611 const MachineOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
612 const MachineOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
614 if ((BaseReg.getReg() != 0 &&
615 X86MCRegisterClasses[X86::GR16RegClassID].contains(BaseReg.getReg())) ||
616 (IndexReg.getReg() != 0 &&
617 X86MCRegisterClasses[X86::GR16RegClassID].contains(IndexReg.getReg())))
622 /// Is32BitMemOperand - Return true if the specified instruction has
623 /// a 32-bit memory operand. Op specifies the operand # of the memoperand.
624 static bool Is32BitMemOperand(const MachineInstr &MI, unsigned Op) {
625 const MachineOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
626 const MachineOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
628 if ((BaseReg.getReg() != 0 &&
629 X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) ||
630 (IndexReg.getReg() != 0 &&
631 X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg())))
636 /// Is64BitMemOperand - Return true if the specified instruction has
637 /// a 64-bit memory operand. Op specifies the operand # of the memoperand.
639 static bool Is64BitMemOperand(const MachineInstr &MI, unsigned Op) {
640 const MachineOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg);
641 const MachineOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg);
643 if ((BaseReg.getReg() != 0 &&
644 X86MCRegisterClasses[X86::GR64RegClassID].contains(BaseReg.getReg())) ||
645 (IndexReg.getReg() != 0 &&
646 X86MCRegisterClasses[X86::GR64RegClassID].contains(IndexReg.getReg())))
652 template<class CodeEmitter>
653 void Emitter<CodeEmitter>::emitOpcodePrefix(uint64_t TSFlags,
655 const MachineInstr &MI,
656 const MCInstrDesc *Desc) const {
657 // Emit the lock opcode prefix as needed.
658 if (Desc->TSFlags & X86II::LOCK)
661 // Emit segment override opcode prefix as needed.
662 emitSegmentOverridePrefix(TSFlags, MemOperand, MI);
664 // Emit the repeat opcode prefix as needed.
665 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP)
668 // Emit the address size opcode prefix as needed.
669 bool need_address_override;
670 if (TSFlags & X86II::AdSize) {
671 need_address_override = true;
672 } else if (MemOperand == -1) {
673 need_address_override = false;
674 } else if (Is64BitMode) {
675 assert(!Is16BitMemOperand(MI, MemOperand));
676 need_address_override = Is32BitMemOperand(MI, MemOperand);
678 assert(!Is64BitMemOperand(MI, MemOperand));
679 need_address_override = Is16BitMemOperand(MI, MemOperand);
682 if (need_address_override)
685 // Emit the operand size opcode prefix as needed.
686 if (TSFlags & X86II::OpSize)
689 bool Need0FPrefix = false;
690 switch (Desc->TSFlags & X86II::Op0Mask) {
691 case X86II::TB: // Two-byte opcode prefix
692 case X86II::T8: // 0F 38
693 case X86II::TA: // 0F 3A
694 case X86II::A6: // 0F A6
695 case X86II::A7: // 0F A7
698 case X86II::REP: break; // already handled.
699 case X86II::T8XS: // F3 0F 38
700 case X86II::XS: // F3 0F
704 case X86II::T8XD: // F2 0F 38
705 case X86II::TAXD: // F2 0F 3A
706 case X86II::XD: // F2 0F
710 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
711 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
713 (((Desc->TSFlags & X86II::Op0Mask)-X86II::D8)
714 >> X86II::Op0Shift));
715 break; // Two-byte opcode prefix
716 default: llvm_unreachable("Invalid prefix!");
717 case 0: break; // No prefix!
720 // Handle REX prefix.
722 if (unsigned REX = determineREX(MI))
723 MCE.emitByte(0x40 | REX);
726 // 0x0F escape code must be emitted just before the opcode.
730 switch (Desc->TSFlags & X86II::Op0Mask) {
731 case X86II::T8XD: // F2 0F 38
732 case X86II::T8XS: // F3 0F 38
733 case X86II::T8: // 0F 38
736 case X86II::TAXD: // F2 0F 38
737 case X86II::TA: // 0F 3A
740 case X86II::A6: // 0F A6
743 case X86II::A7: // 0F A7
749 // On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the range
750 // 0-7 and the difference between the 2 groups is given by the REX prefix.
751 // In the VEX prefix, registers are seen sequencially from 0-15 and encoded
752 // in 1's complement form, example:
754 // ModRM field => XMM9 => 1
755 // VEX.VVVV => XMM9 => ~9
757 // See table 4-35 of Intel AVX Programming Reference for details.
758 template<class CodeEmitter>
760 Emitter<CodeEmitter>::getVEXRegisterEncoding(const MachineInstr &MI,
761 unsigned OpNum) const {
762 unsigned SrcReg = MI.getOperand(OpNum).getReg();
763 unsigned SrcRegNum = getX86RegNum(MI.getOperand(OpNum).getReg());
764 if (X86II::isX86_64ExtendedReg(SrcReg))
767 // The registers represented through VEX_VVVV should
768 // be encoded in 1's complement form.
769 return (~SrcRegNum) & 0xf;
772 /// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed
773 template<class CodeEmitter>
774 void Emitter<CodeEmitter>::emitSegmentOverridePrefix(uint64_t TSFlags,
776 const MachineInstr &MI) const {
777 switch (TSFlags & X86II::SegOvrMask) {
778 default: llvm_unreachable("Invalid segment!");
780 // No segment override, check for explicit one on memory operand.
781 if (MemOperand != -1) { // If the instruction has a memory operand.
782 switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) {
783 default: llvm_unreachable("Unknown segment register!");
785 case X86::CS: MCE.emitByte(0x2E); break;
786 case X86::SS: MCE.emitByte(0x36); break;
787 case X86::DS: MCE.emitByte(0x3E); break;
788 case X86::ES: MCE.emitByte(0x26); break;
789 case X86::FS: MCE.emitByte(0x64); break;
790 case X86::GS: MCE.emitByte(0x65); break;
803 template<class CodeEmitter>
804 void Emitter<CodeEmitter>::emitVEXOpcodePrefix(uint64_t TSFlags,
806 const MachineInstr &MI,
807 const MCInstrDesc *Desc) const {
808 bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
809 bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
810 bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4;
812 // VEX_R: opcode externsion equivalent to REX.R in
813 // 1's complement (inverted) form
815 // 1: Same as REX_R=0 (must be 1 in 32-bit mode)
816 // 0: Same as REX_R=1 (64 bit mode only)
818 unsigned char VEX_R = 0x1;
820 // VEX_X: equivalent to REX.X, only used when a
821 // register is used for index in SIB Byte.
823 // 1: Same as REX.X=0 (must be 1 in 32-bit mode)
824 // 0: Same as REX.X=1 (64-bit mode only)
825 unsigned char VEX_X = 0x1;
829 // 1: Same as REX_B=0 (ignored in 32-bit mode)
830 // 0: Same as REX_B=1 (64 bit mode only)
832 unsigned char VEX_B = 0x1;
834 // VEX_W: opcode specific (use like REX.W, or used for
835 // opcode extension, or ignored, depending on the opcode byte)
836 unsigned char VEX_W = 0;
838 // XOP: Use XOP prefix byte 0x8f instead of VEX.
841 // VEX_5M (VEX m-mmmmm field):
843 // 0b00000: Reserved for future use
844 // 0b00001: implied 0F leading opcode
845 // 0b00010: implied 0F 38 leading opcode bytes
846 // 0b00011: implied 0F 3A leading opcode bytes
847 // 0b00100-0b11111: Reserved for future use
848 // 0b01000: XOP map select - 08h instructions with imm byte
849 // 0b01001: XOP map select - 09h instructions with no imm byte
850 // 0b01010: XOP map select - 0Ah instructions with imm dword
851 unsigned char VEX_5M = 0x1;
853 // VEX_4V (VEX vvvv field): a register specifier
854 // (in 1's complement form) or 1111 if unused.
855 unsigned char VEX_4V = 0xf;
857 // VEX_L (Vector Length):
859 // 0: scalar or 128-bit vector
862 unsigned char VEX_L = 0;
864 // VEX_PP: opcode extension providing equivalent
865 // functionality of a SIMD prefix
872 unsigned char VEX_PP = 0;
874 // Encode the operand size opcode prefix as needed.
875 if (TSFlags & X86II::OpSize)
878 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_W)
881 if ((TSFlags >> X86II::VEXShift) & X86II::XOP)
884 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_L)
887 switch (TSFlags & X86II::Op0Mask) {
888 default: llvm_unreachable("Invalid prefix!");
889 case X86II::T8: // 0F 38
892 case X86II::TA: // 0F 3A
895 case X86II::T8XS: // F3 0F 38
899 case X86II::T8XD: // F2 0F 38
903 case X86II::TAXD: // F2 0F 3A
907 case X86II::XS: // F3 0F
910 case X86II::XD: // F2 0F
922 case X86II::TB: // VEX_5M/VEX_PP already correct
927 // Classify VEX_B, VEX_4V, VEX_R, VEX_X
928 unsigned NumOps = Desc->getNumOperands();
930 if (NumOps > 1 && Desc->getOperandConstraint(1, MCOI::TIED_TO) == 0)
932 else if (NumOps > 3 && Desc->getOperandConstraint(2, MCOI::TIED_TO) == 0) {
933 assert(Desc->getOperandConstraint(NumOps - 1, MCOI::TIED_TO) == 1);
934 // Special case for GATHER with 2 TIED_TO operands
935 // Skip the first 2 operands: dst, mask_wb
939 switch (TSFlags & X86II::FormMask) {
940 case X86II::MRMDestMem: {
941 // MRMDestMem instructions forms:
942 // MemAddr, src1(ModR/M)
943 // MemAddr, src1(VEX_4V), src2(ModR/M)
944 // MemAddr, src1(ModR/M), imm8
946 if (X86II::isX86_64ExtendedReg(MI.getOperand(X86::AddrBaseReg).getReg()))
948 if (X86II::isX86_64ExtendedReg(MI.getOperand(X86::AddrIndexReg).getReg()))
951 CurOp = X86::AddrNumOperands;
953 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
955 const MachineOperand &MO = MI.getOperand(CurOp);
956 if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg()))
960 case X86II::MRMSrcMem:
961 // MRMSrcMem instructions forms:
962 // src1(ModR/M), MemAddr
963 // src1(ModR/M), src2(VEX_4V), MemAddr
964 // src1(ModR/M), MemAddr, imm8
965 // src1(ModR/M), MemAddr, src2(VEX_I8IMM)
968 // dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
969 // dst(ModR/M.reg), src1(VEX_4V), src2(VEX_I8IMM), src3(ModR/M),
970 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
975 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
979 if (X86II::isX86_64ExtendedReg(
980 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
982 if (X86II::isX86_64ExtendedReg(
983 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
987 VEX_4V = getVEXRegisterEncoding(MI, CurOp+X86::AddrNumOperands);
989 case X86II::MRM0m: case X86II::MRM1m:
990 case X86II::MRM2m: case X86II::MRM3m:
991 case X86II::MRM4m: case X86II::MRM5m:
992 case X86II::MRM6m: case X86II::MRM7m: {
993 // MRM[0-9]m instructions forms:
995 // src1(VEX_4V), MemAddr
997 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
999 if (X86II::isX86_64ExtendedReg(
1000 MI.getOperand(MemOperand+X86::AddrBaseReg).getReg()))
1002 if (X86II::isX86_64ExtendedReg(
1003 MI.getOperand(MemOperand+X86::AddrIndexReg).getReg()))
1007 case X86II::MRMSrcReg:
1008 // MRMSrcReg instructions forms:
1009 // dst(ModR/M), src1(VEX_4V), src2(ModR/M), src3(VEX_I8IMM)
1010 // dst(ModR/M), src1(ModR/M)
1011 // dst(ModR/M), src1(ModR/M), imm8
1013 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
1018 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
1020 if (HasMemOp4) // Skip second register source (encoded in I8IMM)
1023 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
1027 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
1029 case X86II::MRMDestReg:
1030 // MRMDestReg instructions forms:
1031 // dst(ModR/M), src(ModR/M)
1032 // dst(ModR/M), src(ModR/M), imm8
1033 // dst(ModR/M), src1(VEX_4V), src2(ModR/M)
1034 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
1039 VEX_4V = getVEXRegisterEncoding(MI, CurOp++);
1041 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
1044 case X86II::MRM0r: case X86II::MRM1r:
1045 case X86II::MRM2r: case X86II::MRM3r:
1046 case X86II::MRM4r: case X86II::MRM5r:
1047 case X86II::MRM6r: case X86II::MRM7r:
1048 // MRM0r-MRM7r instructions forms:
1049 // dst(VEX_4V), src(ModR/M), imm8
1050 VEX_4V = getVEXRegisterEncoding(MI, CurOp);
1053 if (X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg()))
1060 // Emit segment override opcode prefix as needed.
1061 emitSegmentOverridePrefix(TSFlags, MemOperand, MI);
1063 // VEX opcode prefix can have 2 or 3 bytes
1066 // +-----+ +--------------+ +-------------------+
1067 // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp |
1068 // +-----+ +--------------+ +-------------------+
1070 // +-----+ +-------------------+
1071 // | C5h | | R | vvvv | L | pp |
1072 // +-----+ +-------------------+
1074 unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3);
1076 if (VEX_B && VEX_X && !VEX_W && !XOP && (VEX_5M == 1)) { // 2 byte VEX prefix
1078 MCE.emitByte(LastByte | (VEX_R << 7));
1082 // 3 byte VEX prefix
1083 MCE.emitByte(XOP ? 0x8F : 0xC4);
1084 MCE.emitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M);
1085 MCE.emitByte(LastByte | (VEX_W << 7));
1088 template<class CodeEmitter>
1089 void Emitter<CodeEmitter>::emitInstruction(MachineInstr &MI,
1090 const MCInstrDesc *Desc) {
1091 DEBUG(dbgs() << MI);
1093 // If this is a pseudo instruction, lower it.
1094 switch (Desc->getOpcode()) {
1095 case X86::ADD16rr_DB: Desc = UpdateOp(MI, II, X86::OR16rr); break;
1096 case X86::ADD32rr_DB: Desc = UpdateOp(MI, II, X86::OR32rr); break;
1097 case X86::ADD64rr_DB: Desc = UpdateOp(MI, II, X86::OR64rr); break;
1098 case X86::ADD16ri_DB: Desc = UpdateOp(MI, II, X86::OR16ri); break;
1099 case X86::ADD32ri_DB: Desc = UpdateOp(MI, II, X86::OR32ri); break;
1100 case X86::ADD64ri32_DB: Desc = UpdateOp(MI, II, X86::OR64ri32); break;
1101 case X86::ADD16ri8_DB: Desc = UpdateOp(MI, II, X86::OR16ri8); break;
1102 case X86::ADD32ri8_DB: Desc = UpdateOp(MI, II, X86::OR32ri8); break;
1103 case X86::ADD64ri8_DB: Desc = UpdateOp(MI, II, X86::OR64ri8); break;
1104 case X86::ACQUIRE_MOV8rm: Desc = UpdateOp(MI, II, X86::MOV8rm); break;
1105 case X86::ACQUIRE_MOV16rm: Desc = UpdateOp(MI, II, X86::MOV16rm); break;
1106 case X86::ACQUIRE_MOV32rm: Desc = UpdateOp(MI, II, X86::MOV32rm); break;
1107 case X86::ACQUIRE_MOV64rm: Desc = UpdateOp(MI, II, X86::MOV64rm); break;
1108 case X86::RELEASE_MOV8mr: Desc = UpdateOp(MI, II, X86::MOV8mr); break;
1109 case X86::RELEASE_MOV16mr: Desc = UpdateOp(MI, II, X86::MOV16mr); break;
1110 case X86::RELEASE_MOV32mr: Desc = UpdateOp(MI, II, X86::MOV32mr); break;
1111 case X86::RELEASE_MOV64mr: Desc = UpdateOp(MI, II, X86::MOV64mr); break;
1115 MCE.processDebugLoc(MI.getDebugLoc(), true);
1117 unsigned Opcode = Desc->Opcode;
1119 // If this is a two-address instruction, skip one of the register operands.
1120 unsigned NumOps = Desc->getNumOperands();
1122 if (NumOps > 1 && Desc->getOperandConstraint(1, MCOI::TIED_TO) == 0)
1124 else if (NumOps > 3 && Desc->getOperandConstraint(2, MCOI::TIED_TO) == 0) {
1125 assert(Desc->getOperandConstraint(NumOps - 1, MCOI::TIED_TO) == 1);
1126 // Special case for GATHER with 2 TIED_TO operands
1127 // Skip the first 2 operands: dst, mask_wb
1131 uint64_t TSFlags = Desc->TSFlags;
1133 // Is this instruction encoded using the AVX VEX prefix?
1134 bool HasVEXPrefix = (TSFlags >> X86II::VEXShift) & X86II::VEX;
1135 // It uses the VEX.VVVV field?
1136 bool HasVEX_4V = (TSFlags >> X86II::VEXShift) & X86II::VEX_4V;
1137 bool HasVEX_4VOp3 = (TSFlags >> X86II::VEXShift) & X86II::VEX_4VOp3;
1138 bool HasMemOp4 = (TSFlags >> X86II::VEXShift) & X86II::MemOp4;
1139 const unsigned MemOp4_I8IMMOperand = 2;
1141 // Determine where the memory operand starts, if present.
1142 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags, Opcode);
1143 if (MemoryOperand != -1) MemoryOperand += CurOp;
1146 emitOpcodePrefix(TSFlags, MemoryOperand, MI, Desc);
1148 emitVEXOpcodePrefix(TSFlags, MemoryOperand, MI, Desc);
1150 unsigned char BaseOpcode = X86II::getBaseOpcodeFor(Desc->TSFlags);
1151 switch (TSFlags & X86II::FormMask) {
1153 llvm_unreachable("Unknown FormMask value in X86 MachineCodeEmitter!");
1155 // Remember the current PC offset, this is the PIC relocation
1159 llvm_unreachable("pseudo instructions should be removed before code"
1161 // Do nothing for Int_MemBarrier - it's just a comment. Add a debug
1162 // to make it slightly easier to see.
1163 case X86::Int_MemBarrier:
1164 DEBUG(dbgs() << "#MEMBARRIER\n");
1167 case TargetOpcode::INLINEASM:
1168 // We allow inline assembler nodes with empty bodies - they can
1169 // implicitly define registers, which is ok for JIT.
1170 if (MI.getOperand(0).getSymbolName()[0])
1171 report_fatal_error("JIT does not support inline asm!");
1173 case TargetOpcode::PROLOG_LABEL:
1174 case TargetOpcode::GC_LABEL:
1175 case TargetOpcode::EH_LABEL:
1176 MCE.emitLabel(MI.getOperand(0).getMCSymbol());
1179 case TargetOpcode::IMPLICIT_DEF:
1180 case TargetOpcode::KILL:
1182 case X86::MOVPC32r: {
1183 // This emits the "call" portion of this pseudo instruction.
1184 MCE.emitByte(BaseOpcode);
1185 emitConstant(0, X86II::getSizeOfImm(Desc->TSFlags));
1186 // Remember PIC base.
1187 PICBaseOffset = (intptr_t) MCE.getCurrentPCOffset();
1188 X86JITInfo *JTI = TM.getJITInfo();
1189 JTI->setPICBase(MCE.getCurrentPCValue());
1195 case X86II::RawFrm: {
1196 MCE.emitByte(BaseOpcode);
1198 if (CurOp == NumOps)
1201 const MachineOperand &MO = MI.getOperand(CurOp++);
1203 DEBUG(dbgs() << "RawFrm CurOp " << CurOp << "\n");
1204 DEBUG(dbgs() << "isMBB " << MO.isMBB() << "\n");
1205 DEBUG(dbgs() << "isGlobal " << MO.isGlobal() << "\n");
1206 DEBUG(dbgs() << "isSymbol " << MO.isSymbol() << "\n");
1207 DEBUG(dbgs() << "isImm " << MO.isImm() << "\n");
1210 emitPCRelativeBlockAddress(MO.getMBB());
1214 if (MO.isGlobal()) {
1215 emitGlobalAddress(MO.getGlobal(), X86::reloc_pcrel_word,
1220 if (MO.isSymbol()) {
1221 emitExternalSymbolAddress(MO.getSymbolName(), X86::reloc_pcrel_word);
1225 // FIXME: Only used by hackish MCCodeEmitter, remove when dead.
1227 emitJumpTableAddress(MO.getIndex(), X86::reloc_pcrel_word);
1231 assert(MO.isImm() && "Unknown RawFrm operand!");
1232 if (Opcode == X86::CALLpcrel32 || Opcode == X86::CALL64pcrel32) {
1233 // Fix up immediate operand for pc relative calls.
1234 intptr_t Imm = (intptr_t)MO.getImm();
1235 Imm = Imm - MCE.getCurrentPCValue() - 4;
1236 emitConstant(Imm, X86II::getSizeOfImm(Desc->TSFlags));
1238 emitConstant(MO.getImm(), X86II::getSizeOfImm(Desc->TSFlags));
1242 case X86II::AddRegFrm: {
1243 MCE.emitByte(BaseOpcode +
1244 getX86RegNum(MI.getOperand(CurOp++).getReg()));
1246 if (CurOp == NumOps)
1249 const MachineOperand &MO1 = MI.getOperand(CurOp++);
1250 unsigned Size = X86II::getSizeOfImm(Desc->TSFlags);
1252 emitConstant(MO1.getImm(), Size);
1256 unsigned rt = Is64BitMode ? X86::reloc_pcrel_word
1257 : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word);
1258 if (Opcode == X86::MOV32ri64)
1259 rt = X86::reloc_absolute_word; // FIXME: add X86II flag?
1260 // This should not occur on Darwin for relocatable objects.
1261 if (Opcode == X86::MOV64ri)
1262 rt = X86::reloc_absolute_dword; // FIXME: add X86II flag?
1263 if (MO1.isGlobal()) {
1264 bool Indirect = gvNeedsNonLazyPtr(MO1, TM);
1265 emitGlobalAddress(MO1.getGlobal(), rt, MO1.getOffset(), 0,
1267 } else if (MO1.isSymbol())
1268 emitExternalSymbolAddress(MO1.getSymbolName(), rt);
1269 else if (MO1.isCPI())
1270 emitConstPoolAddress(MO1.getIndex(), rt);
1271 else if (MO1.isJTI())
1272 emitJumpTableAddress(MO1.getIndex(), rt);
1276 case X86II::MRMDestReg: {
1277 MCE.emitByte(BaseOpcode);
1279 unsigned SrcRegNum = CurOp+1;
1280 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1283 emitRegModRMByte(MI.getOperand(CurOp).getReg(),
1284 getX86RegNum(MI.getOperand(SrcRegNum).getReg()));
1285 CurOp = SrcRegNum + 1;
1288 case X86II::MRMDestMem: {
1289 MCE.emitByte(BaseOpcode);
1291 unsigned SrcRegNum = CurOp + X86::AddrNumOperands;
1292 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1294 emitMemModRMByte(MI, CurOp,
1295 getX86RegNum(MI.getOperand(SrcRegNum).getReg()));
1296 CurOp = SrcRegNum + 1;
1300 case X86II::MRMSrcReg: {
1301 MCE.emitByte(BaseOpcode);
1303 unsigned SrcRegNum = CurOp+1;
1304 if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
1307 if (HasMemOp4) // Skip 2nd src (which is encoded in I8IMM)
1310 emitRegModRMByte(MI.getOperand(SrcRegNum).getReg(),
1311 getX86RegNum(MI.getOperand(CurOp).getReg()));
1312 // 2 operands skipped with HasMemOp4, compensate accordingly
1313 CurOp = HasMemOp4 ? SrcRegNum : SrcRegNum + 1;
1318 case X86II::MRMSrcMem: {
1319 int AddrOperands = X86::AddrNumOperands;
1320 unsigned FirstMemOp = CurOp+1;
1323 ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV).
1325 if (HasMemOp4) // Skip second register source (encoded in I8IMM)
1328 MCE.emitByte(BaseOpcode);
1330 intptr_t PCAdj = (CurOp + AddrOperands + 1 != NumOps) ?
1331 X86II::getSizeOfImm(Desc->TSFlags) : 0;
1332 emitMemModRMByte(MI, FirstMemOp,
1333 getX86RegNum(MI.getOperand(CurOp).getReg()),PCAdj);
1334 CurOp += AddrOperands + 1;
1340 case X86II::MRM0r: case X86II::MRM1r:
1341 case X86II::MRM2r: case X86II::MRM3r:
1342 case X86II::MRM4r: case X86II::MRM5r:
1343 case X86II::MRM6r: case X86II::MRM7r: {
1344 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1346 MCE.emitByte(BaseOpcode);
1347 emitRegModRMByte(MI.getOperand(CurOp++).getReg(),
1348 (Desc->TSFlags & X86II::FormMask)-X86II::MRM0r);
1350 if (CurOp == NumOps)
1353 const MachineOperand &MO1 = MI.getOperand(CurOp++);
1354 unsigned Size = X86II::getSizeOfImm(Desc->TSFlags);
1356 emitConstant(MO1.getImm(), Size);
1360 unsigned rt = Is64BitMode ? X86::reloc_pcrel_word
1361 : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word);
1362 if (Opcode == X86::MOV64ri32)
1363 rt = X86::reloc_absolute_word_sext; // FIXME: add X86II flag?
1364 if (MO1.isGlobal()) {
1365 bool Indirect = gvNeedsNonLazyPtr(MO1, TM);
1366 emitGlobalAddress(MO1.getGlobal(), rt, MO1.getOffset(), 0,
1368 } else if (MO1.isSymbol())
1369 emitExternalSymbolAddress(MO1.getSymbolName(), rt);
1370 else if (MO1.isCPI())
1371 emitConstPoolAddress(MO1.getIndex(), rt);
1372 else if (MO1.isJTI())
1373 emitJumpTableAddress(MO1.getIndex(), rt);
1377 case X86II::MRM0m: case X86II::MRM1m:
1378 case X86II::MRM2m: case X86II::MRM3m:
1379 case X86II::MRM4m: case X86II::MRM5m:
1380 case X86II::MRM6m: case X86II::MRM7m: {
1381 if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV).
1383 intptr_t PCAdj = (CurOp + X86::AddrNumOperands != NumOps) ?
1384 (MI.getOperand(CurOp+X86::AddrNumOperands).isImm() ?
1385 X86II::getSizeOfImm(Desc->TSFlags) : 4) : 0;
1387 MCE.emitByte(BaseOpcode);
1388 emitMemModRMByte(MI, CurOp, (Desc->TSFlags & X86II::FormMask)-X86II::MRM0m,
1390 CurOp += X86::AddrNumOperands;
1392 if (CurOp == NumOps)
1395 const MachineOperand &MO = MI.getOperand(CurOp++);
1396 unsigned Size = X86II::getSizeOfImm(Desc->TSFlags);
1398 emitConstant(MO.getImm(), Size);
1402 unsigned rt = Is64BitMode ? X86::reloc_pcrel_word
1403 : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word);
1404 if (Opcode == X86::MOV64mi32)
1405 rt = X86::reloc_absolute_word_sext; // FIXME: add X86II flag?
1406 if (MO.isGlobal()) {
1407 bool Indirect = gvNeedsNonLazyPtr(MO, TM);
1408 emitGlobalAddress(MO.getGlobal(), rt, MO.getOffset(), 0,
1410 } else if (MO.isSymbol())
1411 emitExternalSymbolAddress(MO.getSymbolName(), rt);
1412 else if (MO.isCPI())
1413 emitConstPoolAddress(MO.getIndex(), rt);
1414 else if (MO.isJTI())
1415 emitJumpTableAddress(MO.getIndex(), rt);
1419 case X86II::MRM_C1: case X86II::MRM_C2: case X86II::MRM_C3:
1420 case X86II::MRM_C4: case X86II::MRM_C8: case X86II::MRM_C9:
1421 case X86II::MRM_CA: case X86II::MRM_CB: case X86II::MRM_D0:
1422 case X86II::MRM_D1: case X86II::MRM_D4: case X86II::MRM_D5:
1423 case X86II::MRM_D6: case X86II::MRM_D8: case X86II::MRM_D9:
1424 case X86II::MRM_DA: case X86II::MRM_DB: case X86II::MRM_DC:
1425 case X86II::MRM_DD: case X86II::MRM_DE: case X86II::MRM_DF:
1426 case X86II::MRM_E8: case X86II::MRM_F0: case X86II::MRM_F8:
1427 MCE.emitByte(BaseOpcode);
1430 switch (TSFlags & X86II::FormMask) {
1431 default: llvm_unreachable("Invalid Form");
1432 case X86II::MRM_C1: MRM = 0xC1; break;
1433 case X86II::MRM_C2: MRM = 0xC2; break;
1434 case X86II::MRM_C3: MRM = 0xC3; break;
1435 case X86II::MRM_C4: MRM = 0xC4; break;
1436 case X86II::MRM_C8: MRM = 0xC8; break;
1437 case X86II::MRM_C9: MRM = 0xC9; break;
1438 case X86II::MRM_CA: MRM = 0xCA; break;
1439 case X86II::MRM_CB: MRM = 0xCB; break;
1440 case X86II::MRM_D0: MRM = 0xD0; break;
1441 case X86II::MRM_D1: MRM = 0xD1; break;
1442 case X86II::MRM_D4: MRM = 0xD4; break;
1443 case X86II::MRM_D5: MRM = 0xD5; break;
1444 case X86II::MRM_D6: MRM = 0xD6; break;
1445 case X86II::MRM_D8: MRM = 0xD8; break;
1446 case X86II::MRM_D9: MRM = 0xD9; break;
1447 case X86II::MRM_DA: MRM = 0xDA; break;
1448 case X86II::MRM_DB: MRM = 0xDB; break;
1449 case X86II::MRM_DC: MRM = 0xDC; break;
1450 case X86II::MRM_DD: MRM = 0xDD; break;
1451 case X86II::MRM_DE: MRM = 0xDE; break;
1452 case X86II::MRM_DF: MRM = 0xDF; break;
1453 case X86II::MRM_E8: MRM = 0xE8; break;
1454 case X86II::MRM_F0: MRM = 0xF0; break;
1455 case X86II::MRM_F8: MRM = 0xF8; break;
1456 case X86II::MRM_F9: MRM = 0xF9; break;
1462 while (CurOp != NumOps && NumOps - CurOp <= 2) {
1463 // The last source register of a 4 operand instruction in AVX is encoded
1464 // in bits[7:4] of a immediate byte.
1465 if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) {
1466 const MachineOperand &MO = MI.getOperand(HasMemOp4 ? MemOp4_I8IMMOperand
1469 unsigned RegNum = getX86RegNum(MO.getReg()) << 4;
1470 if (X86II::isX86_64ExtendedReg(MO.getReg()))
1472 // If there is an additional 5th operand it must be an immediate, which
1473 // is encoded in bits[3:0]
1474 if (CurOp != NumOps) {
1475 const MachineOperand &MIMM = MI.getOperand(CurOp++);
1477 unsigned Val = MIMM.getImm();
1478 assert(Val < 16 && "Immediate operand value out of range");
1482 emitConstant(RegNum, 1);
1484 emitConstant(MI.getOperand(CurOp++).getImm(),
1485 X86II::getSizeOfImm(Desc->TSFlags));
1489 if (!MI.isVariadic() && CurOp != NumOps) {
1491 dbgs() << "Cannot encode all operands of: " << MI << "\n";
1493 llvm_unreachable(0);
1496 MCE.processDebugLoc(MI.getDebugLoc(), false);